4456f31a8f
We should use COPY here for no-imm operands to reduce the number of generated mov in the isa. However, there is an issue in https://github.com/llvm/llvm-project/pull/162389#discussion_r2430459341 here that blocked me from doing it. With https://github.com/llvm/llvm-project/pull/185751 this should work now
47927 lines
2.2 MiB
47927 lines
2.2 MiB
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
|
|
|
|
; RUN: llc -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefix=SI %s
|
|
; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefix=VI %s
|
|
; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9 %s
|
|
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s
|
|
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s
|
|
|
|
define <28 x float> @bitcast_v28i32_to_v28f32(<28 x i32> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v28i32_to_v28f32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB0_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 3, v27
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 3, v25
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: .LBB0_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28i32_to_v28f32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB0_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
|
|
; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
|
|
; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
|
|
; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
|
|
; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
|
|
; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
|
|
; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: .LBB0_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28i32_to_v28f32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB0_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
|
|
; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
|
|
; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
|
|
; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
|
|
; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
|
|
; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
|
|
; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
|
|
; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
|
|
; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
|
|
; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
|
|
; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
|
|
; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
|
|
; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
|
|
; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
|
|
; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
|
|
; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
|
|
; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
|
|
; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
|
|
; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
|
|
; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
|
|
; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
|
|
; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
|
|
; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
|
|
; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
|
|
; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
|
|
; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
|
|
; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
|
|
; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
|
|
; GFX9-NEXT: .LBB0_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v28i32_to_v28f32:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB0_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v27, 3, v27
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v26, 3, v26
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v25, 3, v25
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v24, 3, v24
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v23, 3, v23
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v22, 3, v22
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v21, 3, v21
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v20, 3, v20
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
|
|
; GFX11-NEXT: .LBB0_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <28 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <28 x i32> %a1 to <28 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x i32> %a to <28 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x float> %phi
|
|
}
|
|
|
|
define inreg <28 x float> @bitcast_v28i32_to_v28f32_scalar(<28 x i32> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v28i32_to_v28f32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB1_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB1_3
|
|
; SI-NEXT: .LBB1_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s6, s6, 3
|
|
; SI-NEXT: s_add_i32 s7, s7, 3
|
|
; SI-NEXT: s_add_i32 s8, s8, 3
|
|
; SI-NEXT: s_add_i32 s9, s9, 3
|
|
; SI-NEXT: s_add_i32 s10, s10, 3
|
|
; SI-NEXT: s_add_i32 s11, s11, 3
|
|
; SI-NEXT: s_add_i32 s12, s12, 3
|
|
; SI-NEXT: s_add_i32 s13, s13, 3
|
|
; SI-NEXT: s_add_i32 s14, s14, 3
|
|
; SI-NEXT: s_add_i32 s15, s15, 3
|
|
; SI-NEXT: s_add_i32 s40, s40, 3
|
|
; SI-NEXT: s_add_i32 s41, s41, 3
|
|
; SI-NEXT: s_add_i32 s42, s42, 3
|
|
; SI-NEXT: s_add_i32 s43, s43, 3
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: s_add_i32 s21, s21, 3
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: s_add_i32 s19, s19, 3
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: s_add_i32 s17, s17, 3
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: .LBB1_3: ; %end
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s43
|
|
; SI-NEXT: v_mov_b32_e32 v15, s42
|
|
; SI-NEXT: v_mov_b32_e32 v16, s41
|
|
; SI-NEXT: v_mov_b32_e32 v17, s40
|
|
; SI-NEXT: v_mov_b32_e32 v18, s15
|
|
; SI-NEXT: v_mov_b32_e32 v19, s14
|
|
; SI-NEXT: v_mov_b32_e32 v20, s13
|
|
; SI-NEXT: v_mov_b32_e32 v21, s12
|
|
; SI-NEXT: v_mov_b32_e32 v22, s11
|
|
; SI-NEXT: v_mov_b32_e32 v23, s10
|
|
; SI-NEXT: v_mov_b32_e32 v24, s9
|
|
; SI-NEXT: v_mov_b32_e32 v25, s8
|
|
; SI-NEXT: v_mov_b32_e32 v26, s7
|
|
; SI-NEXT: v_mov_b32_e32 v27, s6
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB1_4:
|
|
; SI-NEXT: s_branch .LBB1_2
|
|
;
|
|
; VI-LABEL: bitcast_v28i32_to_v28f32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB1_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB1_3
|
|
; VI-NEXT: .LBB1_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s6, s6, 3
|
|
; VI-NEXT: s_add_i32 s7, s7, 3
|
|
; VI-NEXT: s_add_i32 s8, s8, 3
|
|
; VI-NEXT: s_add_i32 s9, s9, 3
|
|
; VI-NEXT: s_add_i32 s10, s10, 3
|
|
; VI-NEXT: s_add_i32 s11, s11, 3
|
|
; VI-NEXT: s_add_i32 s12, s12, 3
|
|
; VI-NEXT: s_add_i32 s13, s13, 3
|
|
; VI-NEXT: s_add_i32 s14, s14, 3
|
|
; VI-NEXT: s_add_i32 s15, s15, 3
|
|
; VI-NEXT: s_add_i32 s40, s40, 3
|
|
; VI-NEXT: s_add_i32 s41, s41, 3
|
|
; VI-NEXT: s_add_i32 s42, s42, 3
|
|
; VI-NEXT: s_add_i32 s43, s43, 3
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: .LBB1_3: ; %end
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v14, s43
|
|
; VI-NEXT: v_mov_b32_e32 v15, s42
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB1_4:
|
|
; VI-NEXT: s_branch .LBB1_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v28i32_to_v28f32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB1_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB1_3
|
|
; GFX9-NEXT: .LBB1_2: ; %cmp.true
|
|
; GFX9-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX9-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX9-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX9-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX9-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX9-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX9-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX9-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX9-NEXT: s_add_i32 s14, s14, 3
|
|
; GFX9-NEXT: s_add_i32 s15, s15, 3
|
|
; GFX9-NEXT: s_add_i32 s40, s40, 3
|
|
; GFX9-NEXT: s_add_i32 s41, s41, 3
|
|
; GFX9-NEXT: s_add_i32 s42, s42, 3
|
|
; GFX9-NEXT: s_add_i32 s43, s43, 3
|
|
; GFX9-NEXT: s_add_i32 s29, s29, 3
|
|
; GFX9-NEXT: s_add_i32 s28, s28, 3
|
|
; GFX9-NEXT: s_add_i32 s27, s27, 3
|
|
; GFX9-NEXT: s_add_i32 s26, s26, 3
|
|
; GFX9-NEXT: s_add_i32 s25, s25, 3
|
|
; GFX9-NEXT: s_add_i32 s24, s24, 3
|
|
; GFX9-NEXT: s_add_i32 s23, s23, 3
|
|
; GFX9-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX9-NEXT: s_add_i32 s21, s21, 3
|
|
; GFX9-NEXT: s_add_i32 s20, s20, 3
|
|
; GFX9-NEXT: s_add_i32 s19, s19, 3
|
|
; GFX9-NEXT: s_add_i32 s18, s18, 3
|
|
; GFX9-NEXT: s_add_i32 s17, s17, 3
|
|
; GFX9-NEXT: s_add_i32 s16, s16, 3
|
|
; GFX9-NEXT: .LBB1_3: ; %end
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB1_4:
|
|
; GFX9-NEXT: s_branch .LBB1_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v28i32_to_v28f32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB1_4
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB1_3
|
|
; GFX11-NEXT: .LBB1_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_i32 s4, s4, 3
|
|
; GFX11-NEXT: s_add_i32 s5, s5, 3
|
|
; GFX11-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX11-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX11-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX11-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX11-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX11-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX11-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX11-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX11-NEXT: s_add_i32 s29, s29, 3
|
|
; GFX11-NEXT: s_add_i32 s28, s28, 3
|
|
; GFX11-NEXT: s_add_i32 s27, s27, 3
|
|
; GFX11-NEXT: s_add_i32 s26, s26, 3
|
|
; GFX11-NEXT: s_add_i32 s25, s25, 3
|
|
; GFX11-NEXT: s_add_i32 s24, s24, 3
|
|
; GFX11-NEXT: s_add_i32 s23, s23, 3
|
|
; GFX11-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX11-NEXT: s_add_i32 s21, s21, 3
|
|
; GFX11-NEXT: s_add_i32 s20, s20, 3
|
|
; GFX11-NEXT: s_add_i32 s19, s19, 3
|
|
; GFX11-NEXT: s_add_i32 s18, s18, 3
|
|
; GFX11-NEXT: s_add_i32 s17, s17, 3
|
|
; GFX11-NEXT: s_add_i32 s16, s16, 3
|
|
; GFX11-NEXT: s_add_i32 s3, s3, 3
|
|
; GFX11-NEXT: s_add_i32 s2, s2, 3
|
|
; GFX11-NEXT: s_add_i32 s1, s1, 3
|
|
; GFX11-NEXT: s_add_i32 s0, s0, 3
|
|
; GFX11-NEXT: .LBB1_3: ; %end
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB1_4:
|
|
; GFX11-NEXT: s_branch .LBB1_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <28 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <28 x i32> %a1 to <28 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x i32> %a to <28 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x float> %phi
|
|
}
|
|
|
|
define <28 x i32> @bitcast_v28f32_to_v28i32(<28 x float> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v28f32_to_v28i32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB2_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; SI-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; SI-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; SI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; SI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; SI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; SI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; SI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; SI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; SI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; SI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; SI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; SI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; SI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; SI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; SI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; SI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; SI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; SI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; SI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; SI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; SI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; SI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; SI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; SI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; SI-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; SI-NEXT: .LBB2_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28f32_to_v28i32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB2_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; VI-NEXT: .LBB2_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28f32_to_v28i32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB2_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; GFX9-NEXT: .LBB2_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v28f32_to_v28i32:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB2_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
|
|
; GFX11-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
|
|
; GFX11-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
|
|
; GFX11-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
|
|
; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
|
|
; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
|
|
; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
|
|
; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
|
|
; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
|
|
; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
|
|
; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
|
|
; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
|
|
; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
|
|
; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
|
|
; GFX11-NEXT: .LBB2_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <28 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <28 x float> %a1 to <28 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x float> %a to <28 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x i32> %phi
|
|
}
|
|
|
|
define inreg <28 x i32> @bitcast_v28f32_to_v28i32_scalar(<28 x float> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v28f32_to_v28i32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; SI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; SI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; SI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; SI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; SI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; SI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; SI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; SI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; SI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; SI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; SI-NEXT: s_mov_b32 s49, s29
|
|
; SI-NEXT: s_mov_b32 s48, s28
|
|
; SI-NEXT: s_mov_b32 s47, s27
|
|
; SI-NEXT: s_mov_b32 s46, s26
|
|
; SI-NEXT: s_mov_b32 s45, s25
|
|
; SI-NEXT: s_mov_b32 s44, s24
|
|
; SI-NEXT: s_mov_b32 s43, s23
|
|
; SI-NEXT: s_mov_b32 s42, s22
|
|
; SI-NEXT: s_mov_b32 s41, s21
|
|
; SI-NEXT: s_mov_b32 s40, s20
|
|
; SI-NEXT: s_mov_b32 s39, s19
|
|
; SI-NEXT: s_mov_b32 s38, s18
|
|
; SI-NEXT: s_mov_b32 s37, s17
|
|
; SI-NEXT: s_mov_b32 s36, s16
|
|
; SI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB3_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB3_4
|
|
; SI-NEXT: .LBB3_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; SI-NEXT: s_branch .LBB3_5
|
|
; SI-NEXT: .LBB3_3:
|
|
; SI-NEXT: s_branch .LBB3_2
|
|
; SI-NEXT: .LBB3_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_mov_b32_e32 v28, s64
|
|
; SI-NEXT: v_mov_b32_e32 v29, s65
|
|
; SI-NEXT: v_mov_b32_e32 v30, s66
|
|
; SI-NEXT: v_mov_b32_e32 v31, s67
|
|
; SI-NEXT: .LBB3_5: ; %end
|
|
; SI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; SI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; SI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; SI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; SI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; SI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; SI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; SI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; SI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; SI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; SI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; SI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28f32_to_v28i32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; VI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; VI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; VI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; VI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; VI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; VI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; VI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; VI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; VI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; VI-NEXT: s_mov_b32 s49, s29
|
|
; VI-NEXT: s_mov_b32 s48, s28
|
|
; VI-NEXT: s_mov_b32 s47, s27
|
|
; VI-NEXT: s_mov_b32 s46, s26
|
|
; VI-NEXT: s_mov_b32 s45, s25
|
|
; VI-NEXT: s_mov_b32 s44, s24
|
|
; VI-NEXT: s_mov_b32 s43, s23
|
|
; VI-NEXT: s_mov_b32 s42, s22
|
|
; VI-NEXT: s_mov_b32 s41, s21
|
|
; VI-NEXT: s_mov_b32 s40, s20
|
|
; VI-NEXT: s_mov_b32 s39, s19
|
|
; VI-NEXT: s_mov_b32 s38, s18
|
|
; VI-NEXT: s_mov_b32 s37, s17
|
|
; VI-NEXT: s_mov_b32 s36, s16
|
|
; VI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB3_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB3_4
|
|
; VI-NEXT: .LBB3_2: ; %cmp.true
|
|
; VI-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; VI-NEXT: s_branch .LBB3_5
|
|
; VI-NEXT: .LBB3_3:
|
|
; VI-NEXT: s_branch .LBB3_2
|
|
; VI-NEXT: .LBB3_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_mov_b32_e32 v28, s64
|
|
; VI-NEXT: v_mov_b32_e32 v29, s65
|
|
; VI-NEXT: v_mov_b32_e32 v30, s66
|
|
; VI-NEXT: v_mov_b32_e32 v31, s67
|
|
; VI-NEXT: .LBB3_5: ; %end
|
|
; VI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; VI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; VI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; VI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; VI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; VI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; VI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; VI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; VI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; VI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; VI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; VI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28f32_to_v28i32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_mov_b32 s49, s29
|
|
; GFX9-NEXT: s_mov_b32 s48, s28
|
|
; GFX9-NEXT: s_mov_b32 s47, s27
|
|
; GFX9-NEXT: s_mov_b32 s46, s26
|
|
; GFX9-NEXT: s_mov_b32 s45, s25
|
|
; GFX9-NEXT: s_mov_b32 s44, s24
|
|
; GFX9-NEXT: s_mov_b32 s43, s23
|
|
; GFX9-NEXT: s_mov_b32 s42, s22
|
|
; GFX9-NEXT: s_mov_b32 s41, s21
|
|
; GFX9-NEXT: s_mov_b32 s40, s20
|
|
; GFX9-NEXT: s_mov_b32 s39, s19
|
|
; GFX9-NEXT: s_mov_b32 s38, s18
|
|
; GFX9-NEXT: s_mov_b32 s37, s17
|
|
; GFX9-NEXT: s_mov_b32 s36, s16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s55, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s54, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s53, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s52, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s51, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s50, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB3_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB3_4
|
|
; GFX9-NEXT: .LBB3_2: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; GFX9-NEXT: s_branch .LBB3_5
|
|
; GFX9-NEXT: .LBB3_3:
|
|
; GFX9-NEXT: s_branch .LBB3_2
|
|
; GFX9-NEXT: .LBB3_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB3_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v28f32_to_v28i32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s4, -1
|
|
; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX11-NEXT: s_mov_b32 s36, s0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s0, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s63, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v8
|
|
; GFX11-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s61, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s57, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v2
|
|
; GFX11-NEXT: s_mov_b32 s47, s23
|
|
; GFX11-NEXT: s_mov_b32 s46, s22
|
|
; GFX11-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX11-NEXT: s_mov_b32 s45, s21
|
|
; GFX11-NEXT: s_mov_b32 s44, s20
|
|
; GFX11-NEXT: s_mov_b32 s43, s19
|
|
; GFX11-NEXT: s_mov_b32 s42, s18
|
|
; GFX11-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX11-NEXT: s_mov_b32 s48, s24
|
|
; GFX11-NEXT: s_mov_b32 s41, s17
|
|
; GFX11-NEXT: s_mov_b32 s40, s16
|
|
; GFX11-NEXT: s_mov_b32 s39, s3
|
|
; GFX11-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX11-NEXT: s_mov_b32 s49, s25
|
|
; GFX11-NEXT: s_mov_b32 s38, s2
|
|
; GFX11-NEXT: s_mov_b32 s37, s1
|
|
; GFX11-NEXT: s_cmp_lg_u32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX11-NEXT: s_mov_b32 s50, s26
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX11-NEXT: s_mov_b32 s51, s27
|
|
; GFX11-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX11-NEXT: s_mov_b32 s52, s28
|
|
; GFX11-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX11-NEXT: s_mov_b32 s53, s29
|
|
; GFX11-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s54, v0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX11-NEXT: v_readfirstlane_b32 s55, v1
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB3_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB3_4
|
|
; GFX11-NEXT: .LBB3_2: ; %cmp.true
|
|
; GFX11-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; GFX11-NEXT: s_branch .LBB3_5
|
|
; GFX11-NEXT: .LBB3_3:
|
|
; GFX11-NEXT: s_branch .LBB3_2
|
|
; GFX11-NEXT: .LBB3_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
|
|
; GFX11-NEXT: .LBB3_5: ; %end
|
|
; GFX11-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX11-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX11-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX11-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX11-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX11-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX11-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX11-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX11-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX11-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX11-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX11-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
|
|
; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s0
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <28 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <28 x float> %a1 to <28 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x float> %a to <28 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x i32> %phi
|
|
}
|
|
|
|
define <14 x i64> @bitcast_v28i32_to_v14i64(<28 x i32> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v28i32_to_v14i64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB4_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 3, v27
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 3, v25
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: .LBB4_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28i32_to_v14i64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB4_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
|
|
; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
|
|
; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
|
|
; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
|
|
; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
|
|
; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
|
|
; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: .LBB4_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28i32_to_v14i64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB4_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
|
|
; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
|
|
; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
|
|
; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
|
|
; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
|
|
; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
|
|
; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
|
|
; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
|
|
; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
|
|
; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
|
|
; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
|
|
; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
|
|
; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
|
|
; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
|
|
; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
|
|
; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
|
|
; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
|
|
; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
|
|
; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
|
|
; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
|
|
; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
|
|
; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
|
|
; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
|
|
; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
|
|
; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
|
|
; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
|
|
; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
|
|
; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
|
|
; GFX9-NEXT: .LBB4_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v28i32_to_v14i64:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB4_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v27, 3, v27
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v26, 3, v26
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v25, 3, v25
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v24, 3, v24
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v23, 3, v23
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v22, 3, v22
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v21, 3, v21
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v20, 3, v20
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
|
|
; GFX11-NEXT: .LBB4_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <28 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <28 x i32> %a1 to <14 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x i32> %a to <14 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x i64> %phi
|
|
}
|
|
|
|
define inreg <14 x i64> @bitcast_v28i32_to_v14i64_scalar(<28 x i32> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v28i32_to_v14i64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB5_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB5_3
|
|
; SI-NEXT: .LBB5_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s6, s6, 3
|
|
; SI-NEXT: s_add_i32 s7, s7, 3
|
|
; SI-NEXT: s_add_i32 s8, s8, 3
|
|
; SI-NEXT: s_add_i32 s9, s9, 3
|
|
; SI-NEXT: s_add_i32 s10, s10, 3
|
|
; SI-NEXT: s_add_i32 s11, s11, 3
|
|
; SI-NEXT: s_add_i32 s12, s12, 3
|
|
; SI-NEXT: s_add_i32 s13, s13, 3
|
|
; SI-NEXT: s_add_i32 s14, s14, 3
|
|
; SI-NEXT: s_add_i32 s15, s15, 3
|
|
; SI-NEXT: s_add_i32 s40, s40, 3
|
|
; SI-NEXT: s_add_i32 s41, s41, 3
|
|
; SI-NEXT: s_add_i32 s42, s42, 3
|
|
; SI-NEXT: s_add_i32 s43, s43, 3
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: s_add_i32 s21, s21, 3
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: s_add_i32 s19, s19, 3
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: s_add_i32 s17, s17, 3
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: .LBB5_3: ; %end
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s43
|
|
; SI-NEXT: v_mov_b32_e32 v15, s42
|
|
; SI-NEXT: v_mov_b32_e32 v16, s41
|
|
; SI-NEXT: v_mov_b32_e32 v17, s40
|
|
; SI-NEXT: v_mov_b32_e32 v18, s15
|
|
; SI-NEXT: v_mov_b32_e32 v19, s14
|
|
; SI-NEXT: v_mov_b32_e32 v20, s13
|
|
; SI-NEXT: v_mov_b32_e32 v21, s12
|
|
; SI-NEXT: v_mov_b32_e32 v22, s11
|
|
; SI-NEXT: v_mov_b32_e32 v23, s10
|
|
; SI-NEXT: v_mov_b32_e32 v24, s9
|
|
; SI-NEXT: v_mov_b32_e32 v25, s8
|
|
; SI-NEXT: v_mov_b32_e32 v26, s7
|
|
; SI-NEXT: v_mov_b32_e32 v27, s6
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB5_4:
|
|
; SI-NEXT: s_branch .LBB5_2
|
|
;
|
|
; VI-LABEL: bitcast_v28i32_to_v14i64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB5_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB5_3
|
|
; VI-NEXT: .LBB5_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s6, s6, 3
|
|
; VI-NEXT: s_add_i32 s7, s7, 3
|
|
; VI-NEXT: s_add_i32 s8, s8, 3
|
|
; VI-NEXT: s_add_i32 s9, s9, 3
|
|
; VI-NEXT: s_add_i32 s10, s10, 3
|
|
; VI-NEXT: s_add_i32 s11, s11, 3
|
|
; VI-NEXT: s_add_i32 s12, s12, 3
|
|
; VI-NEXT: s_add_i32 s13, s13, 3
|
|
; VI-NEXT: s_add_i32 s14, s14, 3
|
|
; VI-NEXT: s_add_i32 s15, s15, 3
|
|
; VI-NEXT: s_add_i32 s40, s40, 3
|
|
; VI-NEXT: s_add_i32 s41, s41, 3
|
|
; VI-NEXT: s_add_i32 s42, s42, 3
|
|
; VI-NEXT: s_add_i32 s43, s43, 3
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: .LBB5_3: ; %end
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v14, s43
|
|
; VI-NEXT: v_mov_b32_e32 v15, s42
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB5_4:
|
|
; VI-NEXT: s_branch .LBB5_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v28i32_to_v14i64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB5_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB5_3
|
|
; GFX9-NEXT: .LBB5_2: ; %cmp.true
|
|
; GFX9-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX9-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX9-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX9-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX9-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX9-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX9-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX9-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX9-NEXT: s_add_i32 s14, s14, 3
|
|
; GFX9-NEXT: s_add_i32 s15, s15, 3
|
|
; GFX9-NEXT: s_add_i32 s40, s40, 3
|
|
; GFX9-NEXT: s_add_i32 s41, s41, 3
|
|
; GFX9-NEXT: s_add_i32 s42, s42, 3
|
|
; GFX9-NEXT: s_add_i32 s43, s43, 3
|
|
; GFX9-NEXT: s_add_i32 s29, s29, 3
|
|
; GFX9-NEXT: s_add_i32 s28, s28, 3
|
|
; GFX9-NEXT: s_add_i32 s27, s27, 3
|
|
; GFX9-NEXT: s_add_i32 s26, s26, 3
|
|
; GFX9-NEXT: s_add_i32 s25, s25, 3
|
|
; GFX9-NEXT: s_add_i32 s24, s24, 3
|
|
; GFX9-NEXT: s_add_i32 s23, s23, 3
|
|
; GFX9-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX9-NEXT: s_add_i32 s21, s21, 3
|
|
; GFX9-NEXT: s_add_i32 s20, s20, 3
|
|
; GFX9-NEXT: s_add_i32 s19, s19, 3
|
|
; GFX9-NEXT: s_add_i32 s18, s18, 3
|
|
; GFX9-NEXT: s_add_i32 s17, s17, 3
|
|
; GFX9-NEXT: s_add_i32 s16, s16, 3
|
|
; GFX9-NEXT: .LBB5_3: ; %end
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB5_4:
|
|
; GFX9-NEXT: s_branch .LBB5_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v28i32_to_v14i64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB5_4
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB5_3
|
|
; GFX11-NEXT: .LBB5_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_i32 s4, s4, 3
|
|
; GFX11-NEXT: s_add_i32 s5, s5, 3
|
|
; GFX11-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX11-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX11-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX11-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX11-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX11-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX11-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX11-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX11-NEXT: s_add_i32 s29, s29, 3
|
|
; GFX11-NEXT: s_add_i32 s28, s28, 3
|
|
; GFX11-NEXT: s_add_i32 s27, s27, 3
|
|
; GFX11-NEXT: s_add_i32 s26, s26, 3
|
|
; GFX11-NEXT: s_add_i32 s25, s25, 3
|
|
; GFX11-NEXT: s_add_i32 s24, s24, 3
|
|
; GFX11-NEXT: s_add_i32 s23, s23, 3
|
|
; GFX11-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX11-NEXT: s_add_i32 s21, s21, 3
|
|
; GFX11-NEXT: s_add_i32 s20, s20, 3
|
|
; GFX11-NEXT: s_add_i32 s19, s19, 3
|
|
; GFX11-NEXT: s_add_i32 s18, s18, 3
|
|
; GFX11-NEXT: s_add_i32 s17, s17, 3
|
|
; GFX11-NEXT: s_add_i32 s16, s16, 3
|
|
; GFX11-NEXT: s_add_i32 s3, s3, 3
|
|
; GFX11-NEXT: s_add_i32 s2, s2, 3
|
|
; GFX11-NEXT: s_add_i32 s1, s1, 3
|
|
; GFX11-NEXT: s_add_i32 s0, s0, 3
|
|
; GFX11-NEXT: .LBB5_3: ; %end
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB5_4:
|
|
; GFX11-NEXT: s_branch .LBB5_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <28 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <28 x i32> %a1 to <14 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x i32> %a to <14 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x i64> %phi
|
|
}
|
|
|
|
define <28 x i32> @bitcast_v14i64_to_v28i32(<14 x i64> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v14i64_to_v28i32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB6_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; SI-NEXT: .LBB6_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14i64_to_v28i32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB6_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
|
|
; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
|
|
; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
|
|
; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
|
|
; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
|
|
; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; VI-NEXT: .LBB6_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14i64_to_v28i32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB6_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX9-NEXT: .LBB6_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v14i64_to_v28i32:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB6_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
|
|
; GFX11-NEXT: .LBB6_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <14 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <14 x i64> %a1 to <28 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x i64> %a to <28 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x i32> %phi
|
|
}
|
|
|
|
define inreg <28 x i32> @bitcast_v14i64_to_v28i32_scalar(<14 x i64> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v14i64_to_v28i32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB7_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB7_3
|
|
; SI-NEXT: .LBB7_2: ; %cmp.true
|
|
; SI-NEXT: s_add_u32 s7, s7, 3
|
|
; SI-NEXT: s_addc_u32 s6, s6, 0
|
|
; SI-NEXT: s_add_u32 s9, s9, 3
|
|
; SI-NEXT: s_addc_u32 s8, s8, 0
|
|
; SI-NEXT: s_add_u32 s11, s11, 3
|
|
; SI-NEXT: s_addc_u32 s10, s10, 0
|
|
; SI-NEXT: s_add_u32 s13, s13, 3
|
|
; SI-NEXT: s_addc_u32 s12, s12, 0
|
|
; SI-NEXT: s_add_u32 s15, s15, 3
|
|
; SI-NEXT: s_addc_u32 s14, s14, 0
|
|
; SI-NEXT: s_add_u32 s41, s41, 3
|
|
; SI-NEXT: s_addc_u32 s40, s40, 0
|
|
; SI-NEXT: s_add_u32 s43, s43, 3
|
|
; SI-NEXT: s_addc_u32 s42, s42, 0
|
|
; SI-NEXT: s_add_u32 s28, s28, 3
|
|
; SI-NEXT: s_addc_u32 s29, s29, 0
|
|
; SI-NEXT: s_add_u32 s26, s26, 3
|
|
; SI-NEXT: s_addc_u32 s27, s27, 0
|
|
; SI-NEXT: s_add_u32 s24, s24, 3
|
|
; SI-NEXT: s_addc_u32 s25, s25, 0
|
|
; SI-NEXT: s_add_u32 s22, s22, 3
|
|
; SI-NEXT: s_addc_u32 s23, s23, 0
|
|
; SI-NEXT: s_add_u32 s20, s20, 3
|
|
; SI-NEXT: s_addc_u32 s21, s21, 0
|
|
; SI-NEXT: s_add_u32 s18, s18, 3
|
|
; SI-NEXT: s_addc_u32 s19, s19, 0
|
|
; SI-NEXT: s_add_u32 s16, s16, 3
|
|
; SI-NEXT: s_addc_u32 s17, s17, 0
|
|
; SI-NEXT: .LBB7_3: ; %end
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s43
|
|
; SI-NEXT: v_mov_b32_e32 v15, s42
|
|
; SI-NEXT: v_mov_b32_e32 v16, s41
|
|
; SI-NEXT: v_mov_b32_e32 v17, s40
|
|
; SI-NEXT: v_mov_b32_e32 v18, s15
|
|
; SI-NEXT: v_mov_b32_e32 v19, s14
|
|
; SI-NEXT: v_mov_b32_e32 v20, s13
|
|
; SI-NEXT: v_mov_b32_e32 v21, s12
|
|
; SI-NEXT: v_mov_b32_e32 v22, s11
|
|
; SI-NEXT: v_mov_b32_e32 v23, s10
|
|
; SI-NEXT: v_mov_b32_e32 v24, s9
|
|
; SI-NEXT: v_mov_b32_e32 v25, s8
|
|
; SI-NEXT: v_mov_b32_e32 v26, s7
|
|
; SI-NEXT: v_mov_b32_e32 v27, s6
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB7_4:
|
|
; SI-NEXT: s_branch .LBB7_2
|
|
;
|
|
; VI-LABEL: bitcast_v14i64_to_v28i32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB7_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB7_3
|
|
; VI-NEXT: .LBB7_2: ; %cmp.true
|
|
; VI-NEXT: s_add_u32 s7, s7, 3
|
|
; VI-NEXT: s_addc_u32 s6, s6, 0
|
|
; VI-NEXT: s_add_u32 s9, s9, 3
|
|
; VI-NEXT: s_addc_u32 s8, s8, 0
|
|
; VI-NEXT: s_add_u32 s11, s11, 3
|
|
; VI-NEXT: s_addc_u32 s10, s10, 0
|
|
; VI-NEXT: s_add_u32 s13, s13, 3
|
|
; VI-NEXT: s_addc_u32 s12, s12, 0
|
|
; VI-NEXT: s_add_u32 s15, s15, 3
|
|
; VI-NEXT: s_addc_u32 s14, s14, 0
|
|
; VI-NEXT: s_add_u32 s41, s41, 3
|
|
; VI-NEXT: s_addc_u32 s40, s40, 0
|
|
; VI-NEXT: s_add_u32 s43, s43, 3
|
|
; VI-NEXT: s_addc_u32 s42, s42, 0
|
|
; VI-NEXT: s_add_u32 s28, s28, 3
|
|
; VI-NEXT: s_addc_u32 s29, s29, 0
|
|
; VI-NEXT: s_add_u32 s26, s26, 3
|
|
; VI-NEXT: s_addc_u32 s27, s27, 0
|
|
; VI-NEXT: s_add_u32 s24, s24, 3
|
|
; VI-NEXT: s_addc_u32 s25, s25, 0
|
|
; VI-NEXT: s_add_u32 s22, s22, 3
|
|
; VI-NEXT: s_addc_u32 s23, s23, 0
|
|
; VI-NEXT: s_add_u32 s20, s20, 3
|
|
; VI-NEXT: s_addc_u32 s21, s21, 0
|
|
; VI-NEXT: s_add_u32 s18, s18, 3
|
|
; VI-NEXT: s_addc_u32 s19, s19, 0
|
|
; VI-NEXT: s_add_u32 s16, s16, 3
|
|
; VI-NEXT: s_addc_u32 s17, s17, 0
|
|
; VI-NEXT: .LBB7_3: ; %end
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v14, s43
|
|
; VI-NEXT: v_mov_b32_e32 v15, s42
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB7_4:
|
|
; VI-NEXT: s_branch .LBB7_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v14i64_to_v28i32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB7_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB7_3
|
|
; GFX9-NEXT: .LBB7_2: ; %cmp.true
|
|
; GFX9-NEXT: s_add_u32 s7, s7, 3
|
|
; GFX9-NEXT: s_addc_u32 s6, s6, 0
|
|
; GFX9-NEXT: s_add_u32 s9, s9, 3
|
|
; GFX9-NEXT: s_addc_u32 s8, s8, 0
|
|
; GFX9-NEXT: s_add_u32 s11, s11, 3
|
|
; GFX9-NEXT: s_addc_u32 s10, s10, 0
|
|
; GFX9-NEXT: s_add_u32 s13, s13, 3
|
|
; GFX9-NEXT: s_addc_u32 s12, s12, 0
|
|
; GFX9-NEXT: s_add_u32 s15, s15, 3
|
|
; GFX9-NEXT: s_addc_u32 s14, s14, 0
|
|
; GFX9-NEXT: s_add_u32 s41, s41, 3
|
|
; GFX9-NEXT: s_addc_u32 s40, s40, 0
|
|
; GFX9-NEXT: s_add_u32 s43, s43, 3
|
|
; GFX9-NEXT: s_addc_u32 s42, s42, 0
|
|
; GFX9-NEXT: s_add_u32 s28, s28, 3
|
|
; GFX9-NEXT: s_addc_u32 s29, s29, 0
|
|
; GFX9-NEXT: s_add_u32 s26, s26, 3
|
|
; GFX9-NEXT: s_addc_u32 s27, s27, 0
|
|
; GFX9-NEXT: s_add_u32 s24, s24, 3
|
|
; GFX9-NEXT: s_addc_u32 s25, s25, 0
|
|
; GFX9-NEXT: s_add_u32 s22, s22, 3
|
|
; GFX9-NEXT: s_addc_u32 s23, s23, 0
|
|
; GFX9-NEXT: s_add_u32 s20, s20, 3
|
|
; GFX9-NEXT: s_addc_u32 s21, s21, 0
|
|
; GFX9-NEXT: s_add_u32 s18, s18, 3
|
|
; GFX9-NEXT: s_addc_u32 s19, s19, 0
|
|
; GFX9-NEXT: s_add_u32 s16, s16, 3
|
|
; GFX9-NEXT: s_addc_u32 s17, s17, 0
|
|
; GFX9-NEXT: .LBB7_3: ; %end
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB7_4:
|
|
; GFX9-NEXT: s_branch .LBB7_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v14i64_to_v28i32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB7_4
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB7_3
|
|
; GFX11-NEXT: .LBB7_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_u32 s5, s5, 3
|
|
; GFX11-NEXT: s_addc_u32 s4, s4, 0
|
|
; GFX11-NEXT: s_add_u32 s7, s7, 3
|
|
; GFX11-NEXT: s_addc_u32 s6, s6, 0
|
|
; GFX11-NEXT: s_add_u32 s9, s9, 3
|
|
; GFX11-NEXT: s_addc_u32 s8, s8, 0
|
|
; GFX11-NEXT: s_add_u32 s11, s11, 3
|
|
; GFX11-NEXT: s_addc_u32 s10, s10, 0
|
|
; GFX11-NEXT: s_add_u32 s13, s13, 3
|
|
; GFX11-NEXT: s_addc_u32 s12, s12, 0
|
|
; GFX11-NEXT: s_add_u32 s28, s28, 3
|
|
; GFX11-NEXT: s_addc_u32 s29, s29, 0
|
|
; GFX11-NEXT: s_add_u32 s26, s26, 3
|
|
; GFX11-NEXT: s_addc_u32 s27, s27, 0
|
|
; GFX11-NEXT: s_add_u32 s24, s24, 3
|
|
; GFX11-NEXT: s_addc_u32 s25, s25, 0
|
|
; GFX11-NEXT: s_add_u32 s22, s22, 3
|
|
; GFX11-NEXT: s_addc_u32 s23, s23, 0
|
|
; GFX11-NEXT: s_add_u32 s20, s20, 3
|
|
; GFX11-NEXT: s_addc_u32 s21, s21, 0
|
|
; GFX11-NEXT: s_add_u32 s18, s18, 3
|
|
; GFX11-NEXT: s_addc_u32 s19, s19, 0
|
|
; GFX11-NEXT: s_add_u32 s16, s16, 3
|
|
; GFX11-NEXT: s_addc_u32 s17, s17, 0
|
|
; GFX11-NEXT: s_add_u32 s2, s2, 3
|
|
; GFX11-NEXT: s_addc_u32 s3, s3, 0
|
|
; GFX11-NEXT: s_add_u32 s0, s0, 3
|
|
; GFX11-NEXT: s_addc_u32 s1, s1, 0
|
|
; GFX11-NEXT: .LBB7_3: ; %end
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB7_4:
|
|
; GFX11-NEXT: s_branch .LBB7_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <14 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <14 x i64> %a1 to <28 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x i64> %a to <28 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x i32> %phi
|
|
}
|
|
|
|
define <14 x double> @bitcast_v28i32_to_v14f64(<28 x i32> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v28i32_to_v14f64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB8_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 3, v27
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 3, v25
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: .LBB8_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28i32_to_v14f64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB8_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
|
|
; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
|
|
; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
|
|
; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
|
|
; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
|
|
; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
|
|
; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: .LBB8_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28i32_to_v14f64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB8_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
|
|
; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
|
|
; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
|
|
; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
|
|
; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
|
|
; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
|
|
; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
|
|
; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
|
|
; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
|
|
; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
|
|
; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
|
|
; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
|
|
; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
|
|
; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
|
|
; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
|
|
; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
|
|
; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
|
|
; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
|
|
; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
|
|
; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
|
|
; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
|
|
; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
|
|
; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
|
|
; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
|
|
; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
|
|
; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
|
|
; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
|
|
; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
|
|
; GFX9-NEXT: .LBB8_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v28i32_to_v14f64:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB8_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v27, 3, v27
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v26, 3, v26
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v25, 3, v25
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v24, 3, v24
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v23, 3, v23
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v22, 3, v22
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v21, 3, v21
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v20, 3, v20
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v19, 3, v19
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v18, 3, v18
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v17, 3, v17
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v16, 3, v16
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v15, 3, v15
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v14, 3, v14
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v13, 3, v13
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v12, 3, v12
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v11, 3, v11
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v10, 3, v10
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v9, 3, v9
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v8, 3, v8
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v7, 3, v7
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v6, 3, v6
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v5, 3, v5
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v4, 3, v4
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v3, 3, v3
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v2, 3, v2
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v1, 3, v1
|
|
; GFX11-NEXT: v_add_nc_u32_e32 v0, 3, v0
|
|
; GFX11-NEXT: .LBB8_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <28 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <28 x i32> %a1 to <14 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x i32> %a to <14 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x double> %phi
|
|
}
|
|
|
|
define inreg <14 x double> @bitcast_v28i32_to_v14f64_scalar(<28 x i32> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v28i32_to_v14f64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB9_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB9_3
|
|
; SI-NEXT: .LBB9_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s6, s6, 3
|
|
; SI-NEXT: s_add_i32 s7, s7, 3
|
|
; SI-NEXT: s_add_i32 s8, s8, 3
|
|
; SI-NEXT: s_add_i32 s9, s9, 3
|
|
; SI-NEXT: s_add_i32 s10, s10, 3
|
|
; SI-NEXT: s_add_i32 s11, s11, 3
|
|
; SI-NEXT: s_add_i32 s12, s12, 3
|
|
; SI-NEXT: s_add_i32 s13, s13, 3
|
|
; SI-NEXT: s_add_i32 s14, s14, 3
|
|
; SI-NEXT: s_add_i32 s15, s15, 3
|
|
; SI-NEXT: s_add_i32 s40, s40, 3
|
|
; SI-NEXT: s_add_i32 s41, s41, 3
|
|
; SI-NEXT: s_add_i32 s42, s42, 3
|
|
; SI-NEXT: s_add_i32 s43, s43, 3
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: s_add_i32 s21, s21, 3
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: s_add_i32 s19, s19, 3
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: s_add_i32 s17, s17, 3
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: .LBB9_3: ; %end
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s43
|
|
; SI-NEXT: v_mov_b32_e32 v15, s42
|
|
; SI-NEXT: v_mov_b32_e32 v16, s41
|
|
; SI-NEXT: v_mov_b32_e32 v17, s40
|
|
; SI-NEXT: v_mov_b32_e32 v18, s15
|
|
; SI-NEXT: v_mov_b32_e32 v19, s14
|
|
; SI-NEXT: v_mov_b32_e32 v20, s13
|
|
; SI-NEXT: v_mov_b32_e32 v21, s12
|
|
; SI-NEXT: v_mov_b32_e32 v22, s11
|
|
; SI-NEXT: v_mov_b32_e32 v23, s10
|
|
; SI-NEXT: v_mov_b32_e32 v24, s9
|
|
; SI-NEXT: v_mov_b32_e32 v25, s8
|
|
; SI-NEXT: v_mov_b32_e32 v26, s7
|
|
; SI-NEXT: v_mov_b32_e32 v27, s6
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB9_4:
|
|
; SI-NEXT: s_branch .LBB9_2
|
|
;
|
|
; VI-LABEL: bitcast_v28i32_to_v14f64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB9_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB9_3
|
|
; VI-NEXT: .LBB9_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s6, s6, 3
|
|
; VI-NEXT: s_add_i32 s7, s7, 3
|
|
; VI-NEXT: s_add_i32 s8, s8, 3
|
|
; VI-NEXT: s_add_i32 s9, s9, 3
|
|
; VI-NEXT: s_add_i32 s10, s10, 3
|
|
; VI-NEXT: s_add_i32 s11, s11, 3
|
|
; VI-NEXT: s_add_i32 s12, s12, 3
|
|
; VI-NEXT: s_add_i32 s13, s13, 3
|
|
; VI-NEXT: s_add_i32 s14, s14, 3
|
|
; VI-NEXT: s_add_i32 s15, s15, 3
|
|
; VI-NEXT: s_add_i32 s40, s40, 3
|
|
; VI-NEXT: s_add_i32 s41, s41, 3
|
|
; VI-NEXT: s_add_i32 s42, s42, 3
|
|
; VI-NEXT: s_add_i32 s43, s43, 3
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: .LBB9_3: ; %end
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v14, s43
|
|
; VI-NEXT: v_mov_b32_e32 v15, s42
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB9_4:
|
|
; VI-NEXT: s_branch .LBB9_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v28i32_to_v14f64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB9_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB9_3
|
|
; GFX9-NEXT: .LBB9_2: ; %cmp.true
|
|
; GFX9-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX9-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX9-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX9-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX9-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX9-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX9-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX9-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX9-NEXT: s_add_i32 s14, s14, 3
|
|
; GFX9-NEXT: s_add_i32 s15, s15, 3
|
|
; GFX9-NEXT: s_add_i32 s40, s40, 3
|
|
; GFX9-NEXT: s_add_i32 s41, s41, 3
|
|
; GFX9-NEXT: s_add_i32 s42, s42, 3
|
|
; GFX9-NEXT: s_add_i32 s43, s43, 3
|
|
; GFX9-NEXT: s_add_i32 s29, s29, 3
|
|
; GFX9-NEXT: s_add_i32 s28, s28, 3
|
|
; GFX9-NEXT: s_add_i32 s27, s27, 3
|
|
; GFX9-NEXT: s_add_i32 s26, s26, 3
|
|
; GFX9-NEXT: s_add_i32 s25, s25, 3
|
|
; GFX9-NEXT: s_add_i32 s24, s24, 3
|
|
; GFX9-NEXT: s_add_i32 s23, s23, 3
|
|
; GFX9-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX9-NEXT: s_add_i32 s21, s21, 3
|
|
; GFX9-NEXT: s_add_i32 s20, s20, 3
|
|
; GFX9-NEXT: s_add_i32 s19, s19, 3
|
|
; GFX9-NEXT: s_add_i32 s18, s18, 3
|
|
; GFX9-NEXT: s_add_i32 s17, s17, 3
|
|
; GFX9-NEXT: s_add_i32 s16, s16, 3
|
|
; GFX9-NEXT: .LBB9_3: ; %end
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB9_4:
|
|
; GFX9-NEXT: s_branch .LBB9_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v28i32_to_v14f64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB9_4
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB9_3
|
|
; GFX11-NEXT: .LBB9_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_i32 s4, s4, 3
|
|
; GFX11-NEXT: s_add_i32 s5, s5, 3
|
|
; GFX11-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX11-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX11-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX11-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX11-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX11-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX11-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX11-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX11-NEXT: s_add_i32 s29, s29, 3
|
|
; GFX11-NEXT: s_add_i32 s28, s28, 3
|
|
; GFX11-NEXT: s_add_i32 s27, s27, 3
|
|
; GFX11-NEXT: s_add_i32 s26, s26, 3
|
|
; GFX11-NEXT: s_add_i32 s25, s25, 3
|
|
; GFX11-NEXT: s_add_i32 s24, s24, 3
|
|
; GFX11-NEXT: s_add_i32 s23, s23, 3
|
|
; GFX11-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX11-NEXT: s_add_i32 s21, s21, 3
|
|
; GFX11-NEXT: s_add_i32 s20, s20, 3
|
|
; GFX11-NEXT: s_add_i32 s19, s19, 3
|
|
; GFX11-NEXT: s_add_i32 s18, s18, 3
|
|
; GFX11-NEXT: s_add_i32 s17, s17, 3
|
|
; GFX11-NEXT: s_add_i32 s16, s16, 3
|
|
; GFX11-NEXT: s_add_i32 s3, s3, 3
|
|
; GFX11-NEXT: s_add_i32 s2, s2, 3
|
|
; GFX11-NEXT: s_add_i32 s1, s1, 3
|
|
; GFX11-NEXT: s_add_i32 s0, s0, 3
|
|
; GFX11-NEXT: .LBB9_3: ; %end
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB9_4:
|
|
; GFX11-NEXT: s_branch .LBB9_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <28 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <28 x i32> %a1 to <14 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x i32> %a to <14 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x double> %phi
|
|
}
|
|
|
|
define <28 x i32> @bitcast_v14f64_to_v28i32(<14 x double> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v14f64_to_v28i32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB10_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; SI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; SI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; SI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; SI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; SI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; SI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; SI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; SI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; SI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; SI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; SI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; SI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; SI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; SI-NEXT: .LBB10_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14f64_to_v28i32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB10_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; VI-NEXT: .LBB10_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14f64_to_v28i32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB10_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX9-NEXT: .LBB10_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v14f64_to_v28i32:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB10_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX11-NEXT: .LBB10_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <14 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <14 x double> %a1 to <28 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x double> %a to <28 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x i32> %phi
|
|
}
|
|
|
|
define inreg <28 x i32> @bitcast_v14f64_to_v28i32_scalar(<14 x double> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v14f64_to_v28i32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; SI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; SI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; SI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; SI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; SI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; SI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; SI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; SI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; SI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; SI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; SI-NEXT: s_mov_b32 s49, s29
|
|
; SI-NEXT: s_mov_b32 s48, s28
|
|
; SI-NEXT: s_mov_b32 s47, s27
|
|
; SI-NEXT: s_mov_b32 s46, s26
|
|
; SI-NEXT: s_mov_b32 s45, s25
|
|
; SI-NEXT: s_mov_b32 s44, s24
|
|
; SI-NEXT: s_mov_b32 s43, s23
|
|
; SI-NEXT: s_mov_b32 s42, s22
|
|
; SI-NEXT: s_mov_b32 s41, s21
|
|
; SI-NEXT: s_mov_b32 s40, s20
|
|
; SI-NEXT: s_mov_b32 s39, s19
|
|
; SI-NEXT: s_mov_b32 s38, s18
|
|
; SI-NEXT: s_mov_b32 s37, s17
|
|
; SI-NEXT: s_mov_b32 s36, s16
|
|
; SI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB11_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB11_4
|
|
; SI-NEXT: .LBB11_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; SI-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; SI-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; SI-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; SI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; SI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; SI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; SI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; SI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; SI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; SI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; SI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; SI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; SI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; SI-NEXT: s_branch .LBB11_5
|
|
; SI-NEXT: .LBB11_3:
|
|
; SI-NEXT: s_branch .LBB11_2
|
|
; SI-NEXT: .LBB11_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_mov_b32_e32 v28, s64
|
|
; SI-NEXT: v_mov_b32_e32 v29, s65
|
|
; SI-NEXT: v_mov_b32_e32 v30, s66
|
|
; SI-NEXT: v_mov_b32_e32 v31, s67
|
|
; SI-NEXT: .LBB11_5: ; %end
|
|
; SI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; SI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; SI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; SI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; SI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; SI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; SI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; SI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; SI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; SI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; SI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; SI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14f64_to_v28i32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; VI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; VI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; VI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; VI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; VI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; VI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; VI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; VI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; VI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; VI-NEXT: s_mov_b32 s49, s29
|
|
; VI-NEXT: s_mov_b32 s48, s28
|
|
; VI-NEXT: s_mov_b32 s47, s27
|
|
; VI-NEXT: s_mov_b32 s46, s26
|
|
; VI-NEXT: s_mov_b32 s45, s25
|
|
; VI-NEXT: s_mov_b32 s44, s24
|
|
; VI-NEXT: s_mov_b32 s43, s23
|
|
; VI-NEXT: s_mov_b32 s42, s22
|
|
; VI-NEXT: s_mov_b32 s41, s21
|
|
; VI-NEXT: s_mov_b32 s40, s20
|
|
; VI-NEXT: s_mov_b32 s39, s19
|
|
; VI-NEXT: s_mov_b32 s38, s18
|
|
; VI-NEXT: s_mov_b32 s37, s17
|
|
; VI-NEXT: s_mov_b32 s36, s16
|
|
; VI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB11_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB11_4
|
|
; VI-NEXT: .LBB11_2: ; %cmp.true
|
|
; VI-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; VI-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; VI-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; VI-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; VI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; VI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; VI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; VI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; VI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; VI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; VI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; VI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; VI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; VI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; VI-NEXT: s_branch .LBB11_5
|
|
; VI-NEXT: .LBB11_3:
|
|
; VI-NEXT: s_branch .LBB11_2
|
|
; VI-NEXT: .LBB11_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_mov_b32_e32 v28, s64
|
|
; VI-NEXT: v_mov_b32_e32 v29, s65
|
|
; VI-NEXT: v_mov_b32_e32 v30, s66
|
|
; VI-NEXT: v_mov_b32_e32 v31, s67
|
|
; VI-NEXT: .LBB11_5: ; %end
|
|
; VI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; VI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; VI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; VI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; VI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; VI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; VI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; VI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; VI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; VI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; VI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; VI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14f64_to_v28i32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_mov_b32 s49, s29
|
|
; GFX9-NEXT: s_mov_b32 s48, s28
|
|
; GFX9-NEXT: s_mov_b32 s47, s27
|
|
; GFX9-NEXT: s_mov_b32 s46, s26
|
|
; GFX9-NEXT: s_mov_b32 s45, s25
|
|
; GFX9-NEXT: s_mov_b32 s44, s24
|
|
; GFX9-NEXT: s_mov_b32 s43, s23
|
|
; GFX9-NEXT: s_mov_b32 s42, s22
|
|
; GFX9-NEXT: s_mov_b32 s41, s21
|
|
; GFX9-NEXT: s_mov_b32 s40, s20
|
|
; GFX9-NEXT: s_mov_b32 s39, s19
|
|
; GFX9-NEXT: s_mov_b32 s38, s18
|
|
; GFX9-NEXT: s_mov_b32 s37, s17
|
|
; GFX9-NEXT: s_mov_b32 s36, s16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s55, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s54, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s53, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s52, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s51, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s50, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB11_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB11_4
|
|
; GFX9-NEXT: .LBB11_2: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; GFX9-NEXT: s_branch .LBB11_5
|
|
; GFX9-NEXT: .LBB11_3:
|
|
; GFX9-NEXT: s_branch .LBB11_2
|
|
; GFX9-NEXT: .LBB11_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB11_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v14f64_to_v28i32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s4, -1
|
|
; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX11-NEXT: s_mov_b32 s36, s0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s0, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s63, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v8
|
|
; GFX11-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s61, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s57, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v2
|
|
; GFX11-NEXT: s_mov_b32 s47, s23
|
|
; GFX11-NEXT: s_mov_b32 s46, s22
|
|
; GFX11-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX11-NEXT: s_mov_b32 s45, s21
|
|
; GFX11-NEXT: s_mov_b32 s44, s20
|
|
; GFX11-NEXT: s_mov_b32 s43, s19
|
|
; GFX11-NEXT: s_mov_b32 s42, s18
|
|
; GFX11-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX11-NEXT: s_mov_b32 s48, s24
|
|
; GFX11-NEXT: s_mov_b32 s41, s17
|
|
; GFX11-NEXT: s_mov_b32 s40, s16
|
|
; GFX11-NEXT: s_mov_b32 s39, s3
|
|
; GFX11-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX11-NEXT: s_mov_b32 s49, s25
|
|
; GFX11-NEXT: s_mov_b32 s38, s2
|
|
; GFX11-NEXT: s_mov_b32 s37, s1
|
|
; GFX11-NEXT: s_cmp_lg_u32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX11-NEXT: s_mov_b32 s50, s26
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX11-NEXT: s_mov_b32 s51, s27
|
|
; GFX11-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX11-NEXT: s_mov_b32 s52, s28
|
|
; GFX11-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX11-NEXT: s_mov_b32 s53, s29
|
|
; GFX11-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s54, v0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX11-NEXT: v_readfirstlane_b32 s55, v1
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB11_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB11_4
|
|
; GFX11-NEXT: .LBB11_2: ; %cmp.true
|
|
; GFX11-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; GFX11-NEXT: s_branch .LBB11_5
|
|
; GFX11-NEXT: .LBB11_3:
|
|
; GFX11-NEXT: s_branch .LBB11_2
|
|
; GFX11-NEXT: .LBB11_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
|
|
; GFX11-NEXT: .LBB11_5: ; %end
|
|
; GFX11-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX11-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX11-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX11-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX11-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX11-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX11-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX11-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX11-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX11-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX11-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX11-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
|
|
; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s0
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <14 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <14 x double> %a1 to <28 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x double> %a to <28 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x i32> %phi
|
|
}
|
|
|
|
define <56 x i16> @bitcast_v28i32_to_v56i16(<28 x i32> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v28i32_to_v56i16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB12_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB12_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB12_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 3, v25
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 3, v27
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB12_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v47
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v40
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v48
|
|
; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v41
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v53
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v51
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v48
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v38
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28i32_to_v56i16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB12_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB12_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB12_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
|
|
; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
|
|
; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
|
|
; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
|
|
; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
|
|
; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
|
|
; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB12_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28i32_to_v56i16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB12_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB12_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB12_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
|
|
; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
|
|
; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
|
|
; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
|
|
; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
|
|
; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
|
|
; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
|
|
; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
|
|
; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
|
|
; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
|
|
; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
|
|
; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
|
|
; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
|
|
; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
|
|
; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
|
|
; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
|
|
; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
|
|
; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
|
|
; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
|
|
; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
|
|
; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
|
|
; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
|
|
; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
|
|
; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
|
|
; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
|
|
; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
|
|
; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
|
|
; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB12_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v47, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v46, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v45, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v44, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v43, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v42, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v41, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v40, v7, s4
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v8, v55, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v54, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v53, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v52, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v51, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v49, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v48, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v39, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v38, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v37, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v36, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v35, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v34, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v33, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v32, v23, s4
|
|
; GFX9-NEXT: v_perm_b32 v24, v31, v24, s4
|
|
; GFX9-NEXT: v_perm_b32 v25, v30, v25, s4
|
|
; GFX9-NEXT: v_perm_b32 v26, v29, v26, s4
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s4
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v28i32_to_v56i16:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB12_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v27, 3, v27
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v26, 3, v26
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v25, 3, v25
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v24, 3, v24
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v23, 3, v23
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v22, 3, v22
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v21, 3, v21
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v20, 3, v20
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v19, 3, v19
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v18, 3, v18
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v17, 3, v17
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v16, 3, v16
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v15, 3, v15
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v14, 3, v14
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v13, 3, v13
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v12, 3, v12
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v11, 3, v11
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 3, v10
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 3, v9
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v8
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v7, 3, v7
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v6
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v5
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v4
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 3, v3
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, 3, v2
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v1
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v0
|
|
; GFX11-TRUE16-NEXT: .LBB12_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v28i32_to_v56i16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr67
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB12_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB12_2: ; %Flow
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB12_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v27, 3, v27
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v26, 3, v26
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v25, 3, v25
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v24, 3, v24
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v23, 3, v23
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v22, 3, v22
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v21, 3, v21
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v20, 3, v20
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v19, 3, v19
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v18, 3, v18
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v17, 3, v17
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v16, 3, v16
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v15, 3, v15
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v14, 3, v14
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v13, 3, v13
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v12, 3, v12
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v11, 3, v11
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v10, 3, v10
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 3, v9
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v8
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v7
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 3, v6
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v5
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v4
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v3
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v2
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v1
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB12_4: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v71, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v69, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v68, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v67, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v66, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v65, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v64, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v55, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v54, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v53, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v52, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v51, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v49, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v39, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v37, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v36, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v35, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v34, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v33, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v32, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v31, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v30, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v29, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <28 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <28 x i32> %a1 to <56 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x i32> %a to <56 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x i16> %phi
|
|
}
|
|
|
|
define inreg <56 x i16> @bitcast_v28i32_to_v56i16_scalar(<28 x i32> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v28i32_to_v56i16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v14
|
|
; SI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; SI-NEXT: v_readfirstlane_b32 s5, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s40, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v0
|
|
; SI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; SI-NEXT: s_cbranch_scc0 .LBB13_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s30, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s53, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB13_3
|
|
; SI-NEXT: .LBB13_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s17, s17, 3
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: s_add_i32 s19, s19, 3
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: s_add_i32 s21, s21, 3
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s41, s41, 3
|
|
; SI-NEXT: s_add_i32 s40, s40, 3
|
|
; SI-NEXT: s_add_i32 s15, s15, 3
|
|
; SI-NEXT: s_add_i32 s14, s14, 3
|
|
; SI-NEXT: s_add_i32 s13, s13, 3
|
|
; SI-NEXT: s_add_i32 s12, s12, 3
|
|
; SI-NEXT: s_add_i32 s11, s11, 3
|
|
; SI-NEXT: s_add_i32 s10, s10, 3
|
|
; SI-NEXT: s_add_i32 s9, s9, 3
|
|
; SI-NEXT: s_add_i32 s8, s8, 3
|
|
; SI-NEXT: s_add_i32 s7, s7, 3
|
|
; SI-NEXT: s_add_i32 s6, s6, 3
|
|
; SI-NEXT: s_add_i32 s5, s5, 3
|
|
; SI-NEXT: s_add_i32 s4, s4, 3
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b32 s30, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s53, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: .LBB13_3: ; %end
|
|
; SI-NEXT: s_lshl_b32 s43, s92, 16
|
|
; SI-NEXT: s_and_b32 s16, s16, 0xffff
|
|
; SI-NEXT: s_or_b32 s16, s16, s43
|
|
; SI-NEXT: s_and_b32 s17, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s53, 16
|
|
; SI-NEXT: s_or_b32 s17, s17, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s90, 16
|
|
; SI-NEXT: s_and_b32 s18, s18, 0xffff
|
|
; SI-NEXT: s_or_b32 s18, s18, s43
|
|
; SI-NEXT: s_and_b32 s19, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s52, 16
|
|
; SI-NEXT: s_or_b32 s19, s19, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s88, 16
|
|
; SI-NEXT: s_and_b32 s20, s20, 0xffff
|
|
; SI-NEXT: s_or_b32 s20, s20, s43
|
|
; SI-NEXT: s_and_b32 s21, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s51, 16
|
|
; SI-NEXT: s_or_b32 s21, s21, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s78, 16
|
|
; SI-NEXT: s_and_b32 s22, s22, 0xffff
|
|
; SI-NEXT: s_or_b32 s22, s22, s43
|
|
; SI-NEXT: s_and_b32 s23, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s50, 16
|
|
; SI-NEXT: s_or_b32 s23, s23, s43
|
|
; SI-NEXT: s_and_b32 s24, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s76, 16
|
|
; SI-NEXT: s_or_b32 s24, s24, s43
|
|
; SI-NEXT: s_and_b32 s25, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s49, 16
|
|
; SI-NEXT: s_or_b32 s25, s25, s43
|
|
; SI-NEXT: s_and_b32 s26, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s74, 16
|
|
; SI-NEXT: s_or_b32 s26, s26, s43
|
|
; SI-NEXT: s_and_b32 s27, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s48, 16
|
|
; SI-NEXT: s_or_b32 s27, s27, s43
|
|
; SI-NEXT: s_and_b32 s28, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s72, 16
|
|
; SI-NEXT: s_or_b32 s28, s28, s43
|
|
; SI-NEXT: s_and_b32 s29, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s39, 16
|
|
; SI-NEXT: s_or_b32 s29, s29, s43
|
|
; SI-NEXT: s_and_b32 s40, s40, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s62, 16
|
|
; SI-NEXT: s_or_b32 s40, s40, s43
|
|
; SI-NEXT: s_and_b32 s41, s41, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s38, 16
|
|
; SI-NEXT: s_or_b32 s41, s41, s43
|
|
; SI-NEXT: s_and_b32 s14, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s60, 16
|
|
; SI-NEXT: s_or_b32 s14, s14, s43
|
|
; SI-NEXT: s_and_b32 s15, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s37, 16
|
|
; SI-NEXT: s_or_b32 s15, s15, s43
|
|
; SI-NEXT: s_and_b32 s12, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s58, 16
|
|
; SI-NEXT: s_or_b32 s12, s12, s43
|
|
; SI-NEXT: s_and_b32 s13, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s36, 16
|
|
; SI-NEXT: s_or_b32 s13, s13, s43
|
|
; SI-NEXT: s_and_b32 s10, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s56, 16
|
|
; SI-NEXT: s_or_b32 s10, s10, s43
|
|
; SI-NEXT: s_and_b32 s11, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s35, 16
|
|
; SI-NEXT: s_or_b32 s11, s11, s43
|
|
; SI-NEXT: s_and_b32 s8, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s46, 16
|
|
; SI-NEXT: s_or_b32 s8, s8, s43
|
|
; SI-NEXT: s_and_b32 s9, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s34, 16
|
|
; SI-NEXT: s_or_b32 s9, s9, s43
|
|
; SI-NEXT: s_and_b32 s6, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s44, 16
|
|
; SI-NEXT: s_and_b32 s4, s4, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s42, 16
|
|
; SI-NEXT: s_or_b32 s6, s6, s43
|
|
; SI-NEXT: s_and_b32 s7, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s31, 16
|
|
; SI-NEXT: s_or_b32 s4, s4, s42
|
|
; SI-NEXT: s_and_b32 s5, s5, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s30, 16
|
|
; SI-NEXT: s_or_b32 s7, s7, s43
|
|
; SI-NEXT: s_or_b32 s5, s5, s42
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s40
|
|
; SI-NEXT: v_mov_b32_e32 v15, s41
|
|
; SI-NEXT: v_mov_b32_e32 v16, s14
|
|
; SI-NEXT: v_mov_b32_e32 v17, s15
|
|
; SI-NEXT: v_mov_b32_e32 v18, s12
|
|
; SI-NEXT: v_mov_b32_e32 v19, s13
|
|
; SI-NEXT: v_mov_b32_e32 v20, s10
|
|
; SI-NEXT: v_mov_b32_e32 v21, s11
|
|
; SI-NEXT: v_mov_b32_e32 v22, s8
|
|
; SI-NEXT: v_mov_b32_e32 v23, s9
|
|
; SI-NEXT: v_mov_b32_e32 v24, s6
|
|
; SI-NEXT: v_mov_b32_e32 v25, s7
|
|
; SI-NEXT: v_mov_b32_e32 v26, s4
|
|
; SI-NEXT: v_mov_b32_e32 v27, s5
|
|
; SI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB13_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr92
|
|
; SI-NEXT: ; implicit-def: $sgpr53
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: ; implicit-def: $sgpr52
|
|
; SI-NEXT: ; implicit-def: $sgpr88
|
|
; SI-NEXT: ; implicit-def: $sgpr51
|
|
; SI-NEXT: ; implicit-def: $sgpr78
|
|
; SI-NEXT: ; implicit-def: $sgpr50
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr49
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr48
|
|
; SI-NEXT: ; implicit-def: $sgpr72
|
|
; SI-NEXT: ; implicit-def: $sgpr39
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr38
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr37
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr36
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr35
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr31
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: s_branch .LBB13_2
|
|
;
|
|
; VI-LABEL: bitcast_v28i32_to_v56i16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB13_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s44, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s42, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB13_3
|
|
; VI-NEXT: .LBB13_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s6, s6, 3
|
|
; VI-NEXT: s_add_i32 s7, s7, 3
|
|
; VI-NEXT: s_add_i32 s8, s8, 3
|
|
; VI-NEXT: s_add_i32 s9, s9, 3
|
|
; VI-NEXT: s_add_i32 s10, s10, 3
|
|
; VI-NEXT: s_add_i32 s11, s11, 3
|
|
; VI-NEXT: s_add_i32 s12, s12, 3
|
|
; VI-NEXT: s_add_i32 s13, s13, 3
|
|
; VI-NEXT: s_add_i32 s14, s14, 3
|
|
; VI-NEXT: s_add_i32 s15, s15, 3
|
|
; VI-NEXT: s_add_i32 s40, s40, 3
|
|
; VI-NEXT: s_add_i32 s41, s41, 3
|
|
; VI-NEXT: s_add_i32 s42, s42, 3
|
|
; VI-NEXT: s_add_i32 s43, s43, 3
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: s_lshr_b32 s44, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s42, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: .LBB13_3: ; %end
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s16, s34, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s16
|
|
; VI-NEXT: s_and_b32 s16, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s17, s31, 16
|
|
; VI-NEXT: s_or_b32 s16, s16, s17
|
|
; VI-NEXT: s_and_b32 s17, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s18, s30, 16
|
|
; VI-NEXT: s_or_b32 s17, s17, s18
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s19, s91, 16
|
|
; VI-NEXT: s_or_b32 s18, s18, s19
|
|
; VI-NEXT: s_and_b32 s19, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s20, s90, 16
|
|
; VI-NEXT: s_or_b32 s19, s19, s20
|
|
; VI-NEXT: s_and_b32 s20, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s21, s89, 16
|
|
; VI-NEXT: s_or_b32 s20, s20, s21
|
|
; VI-NEXT: s_and_b32 s21, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s22, s88, 16
|
|
; VI-NEXT: s_or_b32 s21, s21, s22
|
|
; VI-NEXT: s_and_b32 s22, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s23, s79, 16
|
|
; VI-NEXT: s_or_b32 s22, s22, s23
|
|
; VI-NEXT: s_and_b32 s23, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s24, s78, 16
|
|
; VI-NEXT: s_or_b32 s23, s23, s24
|
|
; VI-NEXT: s_and_b32 s24, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s25, s77, 16
|
|
; VI-NEXT: s_or_b32 s24, s24, s25
|
|
; VI-NEXT: s_and_b32 s25, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s26, s76, 16
|
|
; VI-NEXT: s_or_b32 s25, s25, s26
|
|
; VI-NEXT: s_and_b32 s26, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s27, s75, 16
|
|
; VI-NEXT: s_or_b32 s26, s26, s27
|
|
; VI-NEXT: s_and_b32 s27, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s28, s74, 16
|
|
; VI-NEXT: s_or_b32 s27, s27, s28
|
|
; VI-NEXT: s_and_b32 s28, 0xffff, s43
|
|
; VI-NEXT: s_lshl_b32 s29, s73, 16
|
|
; VI-NEXT: s_or_b32 s28, s28, s29
|
|
; VI-NEXT: s_and_b32 s29, 0xffff, s42
|
|
; VI-NEXT: s_lshl_b32 s42, s72, 16
|
|
; VI-NEXT: s_or_b32 s29, s29, s42
|
|
; VI-NEXT: s_and_b32 s41, 0xffff, s41
|
|
; VI-NEXT: s_lshl_b32 s42, s63, 16
|
|
; VI-NEXT: s_or_b32 s41, s41, s42
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s40
|
|
; VI-NEXT: s_lshl_b32 s42, s62, 16
|
|
; VI-NEXT: s_or_b32 s40, s40, s42
|
|
; VI-NEXT: s_and_b32 s15, 0xffff, s15
|
|
; VI-NEXT: s_lshl_b32 s42, s61, 16
|
|
; VI-NEXT: s_or_b32 s15, s15, s42
|
|
; VI-NEXT: s_and_b32 s14, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s42, s60, 16
|
|
; VI-NEXT: s_or_b32 s14, s14, s42
|
|
; VI-NEXT: s_and_b32 s13, 0xffff, s13
|
|
; VI-NEXT: s_lshl_b32 s42, s59, 16
|
|
; VI-NEXT: s_or_b32 s13, s13, s42
|
|
; VI-NEXT: s_and_b32 s12, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s42, s58, 16
|
|
; VI-NEXT: s_or_b32 s12, s12, s42
|
|
; VI-NEXT: s_and_b32 s11, 0xffff, s11
|
|
; VI-NEXT: s_lshl_b32 s42, s57, 16
|
|
; VI-NEXT: s_or_b32 s11, s11, s42
|
|
; VI-NEXT: s_and_b32 s10, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s42, s56, 16
|
|
; VI-NEXT: s_or_b32 s10, s10, s42
|
|
; VI-NEXT: s_and_b32 s9, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s42, s47, 16
|
|
; VI-NEXT: s_or_b32 s9, s9, s42
|
|
; VI-NEXT: s_and_b32 s8, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s42, s46, 16
|
|
; VI-NEXT: s_or_b32 s8, s8, s42
|
|
; VI-NEXT: s_and_b32 s7, 0xffff, s7
|
|
; VI-NEXT: s_lshl_b32 s42, s45, 16
|
|
; VI-NEXT: s_or_b32 s7, s7, s42
|
|
; VI-NEXT: s_and_b32 s6, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s42, s44, 16
|
|
; VI-NEXT: s_or_b32 s6, s6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s16
|
|
; VI-NEXT: v_mov_b32_e32 v3, s17
|
|
; VI-NEXT: v_mov_b32_e32 v4, s18
|
|
; VI-NEXT: v_mov_b32_e32 v5, s19
|
|
; VI-NEXT: v_mov_b32_e32 v6, s20
|
|
; VI-NEXT: v_mov_b32_e32 v7, s21
|
|
; VI-NEXT: v_mov_b32_e32 v8, s22
|
|
; VI-NEXT: v_mov_b32_e32 v9, s23
|
|
; VI-NEXT: v_mov_b32_e32 v10, s24
|
|
; VI-NEXT: v_mov_b32_e32 v11, s25
|
|
; VI-NEXT: v_mov_b32_e32 v12, s26
|
|
; VI-NEXT: v_mov_b32_e32 v13, s27
|
|
; VI-NEXT: v_mov_b32_e32 v14, s28
|
|
; VI-NEXT: v_mov_b32_e32 v15, s29
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB13_4:
|
|
; VI-NEXT: ; implicit-def: $sgpr35
|
|
; VI-NEXT: ; implicit-def: $sgpr34
|
|
; VI-NEXT: ; implicit-def: $sgpr31
|
|
; VI-NEXT: ; implicit-def: $sgpr30
|
|
; VI-NEXT: ; implicit-def: $sgpr91
|
|
; VI-NEXT: ; implicit-def: $sgpr90
|
|
; VI-NEXT: ; implicit-def: $sgpr89
|
|
; VI-NEXT: ; implicit-def: $sgpr88
|
|
; VI-NEXT: ; implicit-def: $sgpr79
|
|
; VI-NEXT: ; implicit-def: $sgpr78
|
|
; VI-NEXT: ; implicit-def: $sgpr77
|
|
; VI-NEXT: ; implicit-def: $sgpr76
|
|
; VI-NEXT: ; implicit-def: $sgpr75
|
|
; VI-NEXT: ; implicit-def: $sgpr74
|
|
; VI-NEXT: ; implicit-def: $sgpr73
|
|
; VI-NEXT: ; implicit-def: $sgpr72
|
|
; VI-NEXT: ; implicit-def: $sgpr63
|
|
; VI-NEXT: ; implicit-def: $sgpr62
|
|
; VI-NEXT: ; implicit-def: $sgpr61
|
|
; VI-NEXT: ; implicit-def: $sgpr60
|
|
; VI-NEXT: ; implicit-def: $sgpr59
|
|
; VI-NEXT: ; implicit-def: $sgpr58
|
|
; VI-NEXT: ; implicit-def: $sgpr57
|
|
; VI-NEXT: ; implicit-def: $sgpr56
|
|
; VI-NEXT: ; implicit-def: $sgpr47
|
|
; VI-NEXT: ; implicit-def: $sgpr46
|
|
; VI-NEXT: ; implicit-def: $sgpr45
|
|
; VI-NEXT: ; implicit-def: $sgpr44
|
|
; VI-NEXT: s_branch .LBB13_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v28i32_to_v56i16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB13_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s44, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s42, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s43, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB13_3
|
|
; GFX9-NEXT: .LBB13_2: ; %cmp.true
|
|
; GFX9-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX9-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX9-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX9-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX9-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX9-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX9-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX9-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX9-NEXT: s_add_i32 s14, s14, 3
|
|
; GFX9-NEXT: s_add_i32 s15, s15, 3
|
|
; GFX9-NEXT: s_add_i32 s40, s40, 3
|
|
; GFX9-NEXT: s_add_i32 s41, s41, 3
|
|
; GFX9-NEXT: s_add_i32 s42, s42, 3
|
|
; GFX9-NEXT: s_add_i32 s43, s43, 3
|
|
; GFX9-NEXT: s_add_i32 s29, s29, 3
|
|
; GFX9-NEXT: s_add_i32 s28, s28, 3
|
|
; GFX9-NEXT: s_add_i32 s27, s27, 3
|
|
; GFX9-NEXT: s_add_i32 s26, s26, 3
|
|
; GFX9-NEXT: s_add_i32 s25, s25, 3
|
|
; GFX9-NEXT: s_add_i32 s24, s24, 3
|
|
; GFX9-NEXT: s_add_i32 s23, s23, 3
|
|
; GFX9-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX9-NEXT: s_add_i32 s21, s21, 3
|
|
; GFX9-NEXT: s_add_i32 s20, s20, 3
|
|
; GFX9-NEXT: s_add_i32 s19, s19, 3
|
|
; GFX9-NEXT: s_add_i32 s18, s18, 3
|
|
; GFX9-NEXT: s_add_i32 s17, s17, 3
|
|
; GFX9-NEXT: s_add_i32 s16, s16, 3
|
|
; GFX9-NEXT: s_lshr_b32 s44, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s42, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s43, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: .LBB13_3: ; %end
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s95
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s94
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s93
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s92
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s91
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s90
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s88
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s28, s43, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s29, s42, s72
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s41, s63
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s40, s62
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s15, s61
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s14, s60
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s13, s59
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s12, s58
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s57
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s56
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s47
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s46
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s45
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s4
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s5
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB13_4:
|
|
; GFX9-NEXT: ; implicit-def: $sgpr95
|
|
; GFX9-NEXT: ; implicit-def: $sgpr94
|
|
; GFX9-NEXT: ; implicit-def: $sgpr93
|
|
; GFX9-NEXT: ; implicit-def: $sgpr92
|
|
; GFX9-NEXT: ; implicit-def: $sgpr91
|
|
; GFX9-NEXT: ; implicit-def: $sgpr90
|
|
; GFX9-NEXT: ; implicit-def: $sgpr89
|
|
; GFX9-NEXT: ; implicit-def: $sgpr88
|
|
; GFX9-NEXT: ; implicit-def: $sgpr79
|
|
; GFX9-NEXT: ; implicit-def: $sgpr78
|
|
; GFX9-NEXT: ; implicit-def: $sgpr77
|
|
; GFX9-NEXT: ; implicit-def: $sgpr76
|
|
; GFX9-NEXT: ; implicit-def: $sgpr75
|
|
; GFX9-NEXT: ; implicit-def: $sgpr74
|
|
; GFX9-NEXT: ; implicit-def: $sgpr73
|
|
; GFX9-NEXT: ; implicit-def: $sgpr72
|
|
; GFX9-NEXT: ; implicit-def: $sgpr63
|
|
; GFX9-NEXT: ; implicit-def: $sgpr62
|
|
; GFX9-NEXT: ; implicit-def: $sgpr61
|
|
; GFX9-NEXT: ; implicit-def: $sgpr60
|
|
; GFX9-NEXT: ; implicit-def: $sgpr59
|
|
; GFX9-NEXT: ; implicit-def: $sgpr58
|
|
; GFX9-NEXT: ; implicit-def: $sgpr57
|
|
; GFX9-NEXT: ; implicit-def: $sgpr56
|
|
; GFX9-NEXT: ; implicit-def: $sgpr47
|
|
; GFX9-NEXT: ; implicit-def: $sgpr46
|
|
; GFX9-NEXT: ; implicit-def: $sgpr45
|
|
; GFX9-NEXT: ; implicit-def: $sgpr44
|
|
; GFX9-NEXT: s_branch .LBB13_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v28i32_to_v56i16_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-NEXT: s_mov_b32 s90, 0
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB13_4
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-NEXT: s_lshr_b32 s14, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s74, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s76, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s77, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s78, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s88, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s0, 16
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s90
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB13_3
|
|
; GFX11-NEXT: .LBB13_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_i32 s4, s4, 3
|
|
; GFX11-NEXT: s_add_i32 s5, s5, 3
|
|
; GFX11-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX11-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX11-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX11-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX11-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX11-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX11-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX11-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX11-NEXT: s_add_i32 s29, s29, 3
|
|
; GFX11-NEXT: s_add_i32 s28, s28, 3
|
|
; GFX11-NEXT: s_add_i32 s27, s27, 3
|
|
; GFX11-NEXT: s_add_i32 s26, s26, 3
|
|
; GFX11-NEXT: s_add_i32 s25, s25, 3
|
|
; GFX11-NEXT: s_add_i32 s24, s24, 3
|
|
; GFX11-NEXT: s_add_i32 s23, s23, 3
|
|
; GFX11-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX11-NEXT: s_add_i32 s21, s21, 3
|
|
; GFX11-NEXT: s_add_i32 s20, s20, 3
|
|
; GFX11-NEXT: s_add_i32 s19, s19, 3
|
|
; GFX11-NEXT: s_add_i32 s18, s18, 3
|
|
; GFX11-NEXT: s_add_i32 s17, s17, 3
|
|
; GFX11-NEXT: s_add_i32 s16, s16, 3
|
|
; GFX11-NEXT: s_add_i32 s3, s3, 3
|
|
; GFX11-NEXT: s_add_i32 s2, s2, 3
|
|
; GFX11-NEXT: s_add_i32 s1, s1, 3
|
|
; GFX11-NEXT: s_add_i32 s0, s0, 3
|
|
; GFX11-NEXT: s_lshr_b32 s14, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s74, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s76, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s77, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s78, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s88, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s0, 16
|
|
; GFX11-NEXT: .LBB13_3: ; %end
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s88
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s78
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s77
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s76
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s74
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s62
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s60
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s59
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s58
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s57
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s56
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s44
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s40
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s14
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB13_4:
|
|
; GFX11-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-NEXT: ; implicit-def: $sgpr14
|
|
; GFX11-NEXT: s_branch .LBB13_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <28 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <28 x i32> %a1 to <56 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x i32> %a to <56 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x i16> %phi
|
|
}
|
|
|
|
define <28 x i32> @bitcast_v56i16_to_v28i32(<56 x i16> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v56i16_to_v28i32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_mov_b32_e32 v48, v19
|
|
; SI-NEXT: v_mov_b32_e32 v49, v18
|
|
; SI-NEXT: v_mov_b32_e32 v50, v17
|
|
; SI-NEXT: v_mov_b32_e32 v51, v16
|
|
; SI-NEXT: v_mov_b32_e32 v52, v15
|
|
; SI-NEXT: v_mov_b32_e32 v53, v14
|
|
; SI-NEXT: v_mov_b32_e32 v54, v13
|
|
; SI-NEXT: v_mov_b32_e32 v55, v12
|
|
; SI-NEXT: v_mov_b32_e32 v40, v11
|
|
; SI-NEXT: v_mov_b32_e32 v41, v10
|
|
; SI-NEXT: v_mov_b32_e32 v42, v9
|
|
; SI-NEXT: v_mov_b32_e32 v43, v8
|
|
; SI-NEXT: v_mov_b32_e32 v44, v7
|
|
; SI-NEXT: v_mov_b32_e32 v45, v6
|
|
; SI-NEXT: v_mov_b32_e32 v46, v5
|
|
; SI-NEXT: v_mov_b32_e32 v47, v4
|
|
; SI-NEXT: v_mov_b32_e32 v56, v3
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_mov_b32_e32 v57, v2
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_mov_b32_e32 v58, v1
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_mov_b32_e32 v59, v0
|
|
; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v26
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v25
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v23
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v22
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v21
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v8, 16, v48
|
|
; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v49
|
|
; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v50
|
|
; SI-NEXT: v_lshrrev_b32_e32 v11, 16, v51
|
|
; SI-NEXT: v_lshrrev_b32_e32 v12, 16, v52
|
|
; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v53
|
|
; SI-NEXT: v_lshrrev_b32_e32 v14, 16, v54
|
|
; SI-NEXT: v_lshrrev_b32_e32 v15, 16, v55
|
|
; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v40
|
|
; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v41
|
|
; SI-NEXT: v_lshrrev_b32_e32 v18, 16, v42
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v43
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v20, 16, v44
|
|
; SI-NEXT: v_lshrrev_b32_e32 v21, 16, v45
|
|
; SI-NEXT: v_lshrrev_b32_e32 v22, 16, v46
|
|
; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v47
|
|
; SI-NEXT: v_lshrrev_b32_e32 v24, 16, v56
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v57
|
|
; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v58
|
|
; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v59
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB14_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v55
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v59
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v58
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v57
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v56
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v47
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v46
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v45
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v44
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v43
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v42
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v41
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v39
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v38
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v62
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v37
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v32
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v36
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v61
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v35
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v60
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v34
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v63
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v33
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v54
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v53
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v52
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v51
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v50
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v49
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v27
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: .LBB14_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB14_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v55
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v59
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v58
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v57
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v56
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v47
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v46
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v45
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v44
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v43
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v42
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v41
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v40
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_or_b32_e32 v0, v39, v0
|
|
; SI-NEXT: s_mov_b32 s6, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v1, v38, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v62, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v37, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v32, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v36, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v61, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v35, v7
|
|
; SI-NEXT: v_or_b32_e32 v8, v60, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v34, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v63, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v33, v11
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v2
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, s6, v3
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v4
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, s6, v5
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v6
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, s6, v7
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v54
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v53
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v52
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v51
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v50
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v49
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v48
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v10
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, s6, v11
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v12
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, s6, v13
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v14
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, s6, v15
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v16
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, s6, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, s6, v19
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, s6, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, s6, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, s6, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 3, v25
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, s6, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 0x30000, v26
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 3, v27
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 0x30000, v27
|
|
; SI-NEXT: .LBB14_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56i16_to_v28i32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_mov_b32_e32 v32, v27
|
|
; VI-NEXT: v_mov_b32_e32 v33, v26
|
|
; VI-NEXT: v_mov_b32_e32 v34, v25
|
|
; VI-NEXT: v_mov_b32_e32 v35, v24
|
|
; VI-NEXT: v_mov_b32_e32 v36, v23
|
|
; VI-NEXT: v_mov_b32_e32 v37, v22
|
|
; VI-NEXT: v_mov_b32_e32 v38, v21
|
|
; VI-NEXT: v_mov_b32_e32 v39, v20
|
|
; VI-NEXT: v_mov_b32_e32 v48, v19
|
|
; VI-NEXT: v_mov_b32_e32 v49, v18
|
|
; VI-NEXT: v_mov_b32_e32 v50, v17
|
|
; VI-NEXT: v_mov_b32_e32 v51, v16
|
|
; VI-NEXT: v_mov_b32_e32 v52, v15
|
|
; VI-NEXT: v_mov_b32_e32 v53, v14
|
|
; VI-NEXT: v_mov_b32_e32 v54, v13
|
|
; VI-NEXT: v_mov_b32_e32 v55, v12
|
|
; VI-NEXT: v_mov_b32_e32 v40, v11
|
|
; VI-NEXT: v_mov_b32_e32 v41, v10
|
|
; VI-NEXT: v_mov_b32_e32 v42, v9
|
|
; VI-NEXT: v_mov_b32_e32 v43, v8
|
|
; VI-NEXT: v_mov_b32_e32 v44, v7
|
|
; VI-NEXT: v_mov_b32_e32 v45, v6
|
|
; VI-NEXT: v_mov_b32_e32 v46, v5
|
|
; VI-NEXT: v_mov_b32_e32 v47, v4
|
|
; VI-NEXT: v_mov_b32_e32 v56, v3
|
|
; VI-NEXT: v_mov_b32_e32 v57, v2
|
|
; VI-NEXT: v_mov_b32_e32 v58, v1
|
|
; VI-NEXT: v_mov_b32_e32 v59, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB14_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_mov_b32_e32 v27, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v27, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v27, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v27, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v27, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v27, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v27, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v27, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v27, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v27, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v27, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v27, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v27, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v27, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v27, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v27, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v27, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v27, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v27, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v27, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v27, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v27, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v24, v27, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v25, v27, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v26, v27, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v27, v27, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v56, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v47, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v46, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v45, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v44, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v43, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v42, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v41, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v40, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v53, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v52, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v51, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v50, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v49, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v39, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v37, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v36, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v35, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v34, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v33, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v32, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: ; implicit-def: $vgpr59
|
|
; VI-NEXT: ; implicit-def: $vgpr58
|
|
; VI-NEXT: ; implicit-def: $vgpr57
|
|
; VI-NEXT: ; implicit-def: $vgpr56
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: .LBB14_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB14_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 3
|
|
; VI-NEXT: v_add_u16_e32 v0, 3, v59
|
|
; VI-NEXT: v_add_u16_sdwa v1, v59, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v58
|
|
; VI-NEXT: v_add_u16_sdwa v3, v58, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v3
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v57
|
|
; VI-NEXT: v_add_u16_sdwa v3, v57, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; VI-NEXT: v_add_u16_e32 v3, 3, v56
|
|
; VI-NEXT: v_add_u16_sdwa v4, v56, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; VI-NEXT: v_add_u16_e32 v4, 3, v47
|
|
; VI-NEXT: v_add_u16_sdwa v5, v47, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; VI-NEXT: v_add_u16_e32 v5, 3, v46
|
|
; VI-NEXT: v_add_u16_sdwa v6, v46, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; VI-NEXT: v_add_u16_e32 v6, 3, v45
|
|
; VI-NEXT: v_add_u16_sdwa v7, v45, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; VI-NEXT: v_add_u16_e32 v7, 3, v44
|
|
; VI-NEXT: v_add_u16_sdwa v8, v44, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; VI-NEXT: v_add_u16_e32 v8, 3, v43
|
|
; VI-NEXT: v_add_u16_sdwa v9, v43, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; VI-NEXT: v_add_u16_e32 v9, 3, v42
|
|
; VI-NEXT: v_add_u16_sdwa v10, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; VI-NEXT: v_add_u16_e32 v10, 3, v41
|
|
; VI-NEXT: v_add_u16_sdwa v11, v41, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v10, v10, v11
|
|
; VI-NEXT: v_add_u16_e32 v11, 3, v40
|
|
; VI-NEXT: v_add_u16_sdwa v12, v40, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; VI-NEXT: v_add_u16_e32 v12, 3, v55
|
|
; VI-NEXT: v_add_u16_sdwa v13, v55, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; VI-NEXT: v_add_u16_e32 v13, 3, v54
|
|
; VI-NEXT: v_add_u16_sdwa v14, v54, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; VI-NEXT: v_add_u16_e32 v14, 3, v53
|
|
; VI-NEXT: v_add_u16_sdwa v15, v53, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; VI-NEXT: v_add_u16_e32 v15, 3, v52
|
|
; VI-NEXT: v_add_u16_sdwa v16, v52, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; VI-NEXT: v_add_u16_e32 v16, 3, v51
|
|
; VI-NEXT: v_add_u16_sdwa v17, v51, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; VI-NEXT: v_add_u16_e32 v17, 3, v50
|
|
; VI-NEXT: v_add_u16_sdwa v18, v50, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; VI-NEXT: v_add_u16_e32 v18, 3, v49
|
|
; VI-NEXT: v_add_u16_sdwa v19, v49, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; VI-NEXT: v_add_u16_e32 v19, 3, v48
|
|
; VI-NEXT: v_add_u16_sdwa v20, v48, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; VI-NEXT: v_add_u16_e32 v20, 3, v39
|
|
; VI-NEXT: v_add_u16_sdwa v21, v39, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; VI-NEXT: v_add_u16_e32 v21, 3, v38
|
|
; VI-NEXT: v_add_u16_sdwa v22, v38, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; VI-NEXT: v_add_u16_e32 v22, 3, v37
|
|
; VI-NEXT: v_add_u16_sdwa v23, v37, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; VI-NEXT: v_add_u16_e32 v23, 3, v36
|
|
; VI-NEXT: v_add_u16_sdwa v24, v36, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; VI-NEXT: v_add_u16_e32 v24, 3, v35
|
|
; VI-NEXT: v_add_u16_sdwa v25, v35, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; VI-NEXT: v_add_u16_e32 v25, 3, v34
|
|
; VI-NEXT: v_add_u16_sdwa v26, v34, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; VI-NEXT: v_add_u16_e32 v26, 3, v33
|
|
; VI-NEXT: v_add_u16_sdwa v28, v33, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; VI-NEXT: v_add_u16_e32 v28, 3, v32
|
|
; VI-NEXT: v_add_u16_sdwa v27, v32, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; VI-NEXT: .LBB14_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56i16_to_v28i32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_mov_b32_e32 v59, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v22
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v13
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v12
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v11
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v10
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v40
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v56, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v57, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v58, v1
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v41
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v47
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v56
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v57
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v58
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v59
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB14_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v37
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: .LBB14_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB14_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB14_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v28i32:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB14_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: .LBB14_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56i16_to_v28i32:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v69, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v71, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v80, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v68, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v67, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v66, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v65, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v64, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v54, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v53, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v52, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v51, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v50, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v49, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v48, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v39, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v38, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v37, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v36, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v35, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v34, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v33, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v32, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v31, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v30, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v29, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB14_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: .LBB14_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <56 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <56 x i16> %a1 to <28 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x i16> %a to <28 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x i32> %phi
|
|
}
|
|
|
|
define inreg <28 x i32> @bitcast_v56i16_to_v28i32_scalar(<56 x i16> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v56i16_to_v28i32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; SI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; SI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; SI-NEXT: v_writelane_b32 v28, s54, 14
|
|
; SI-NEXT: v_writelane_b32 v28, s55, 15
|
|
; SI-NEXT: v_writelane_b32 v28, s64, 16
|
|
; SI-NEXT: v_writelane_b32 v28, s65, 17
|
|
; SI-NEXT: v_writelane_b32 v28, s66, 18
|
|
; SI-NEXT: v_writelane_b32 v28, s67, 19
|
|
; SI-NEXT: v_writelane_b32 v28, s68, 20
|
|
; SI-NEXT: v_writelane_b32 v28, s69, 21
|
|
; SI-NEXT: v_writelane_b32 v28, s70, 22
|
|
; SI-NEXT: v_writelane_b32 v28, s71, 23
|
|
; SI-NEXT: v_writelane_b32 v28, s80, 24
|
|
; SI-NEXT: v_writelane_b32 v28, s81, 25
|
|
; SI-NEXT: v_writelane_b32 v28, s82, 26
|
|
; SI-NEXT: v_writelane_b32 v28, s83, 27
|
|
; SI-NEXT: v_writelane_b32 v28, s84, 28
|
|
; SI-NEXT: v_writelane_b32 v28, s85, 29
|
|
; SI-NEXT: v_writelane_b32 v28, s86, 30
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s73, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s75, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s77, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s88, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s94, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s31, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s69, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s80, v0
|
|
; SI-NEXT: v_writelane_b32 v28, s87, 31
|
|
; SI-NEXT: s_lshr_b32 s78, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s92, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s95, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s68, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s70, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s81, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s82, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s83, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s84, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s85, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s86, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s87, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s6, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s8, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s10, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s12, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s14, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s72, s73, 16
|
|
; SI-NEXT: s_lshr_b32 s74, s75, 16
|
|
; SI-NEXT: s_lshr_b32 s76, s77, 16
|
|
; SI-NEXT: s_lshr_b32 s79, s88, 16
|
|
; SI-NEXT: s_lshr_b32 s90, s91, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s94, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s31, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s69, 16
|
|
; SI-NEXT: s_lshr_b32 s71, s80, 16
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB15_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s36, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s37, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s38, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s39, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s40, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s41, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s42, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s43, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s44, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s34, 16
|
|
; SI-NEXT: s_or_b32 s45, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s95, 16
|
|
; SI-NEXT: s_or_b32 s46, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s92, 16
|
|
; SI-NEXT: s_or_b32 s47, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s48, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; SI-NEXT: s_or_b32 s49, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s80, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s50, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s51, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s31, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s52, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s94, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s53, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; SI-NEXT: s_or_b32 s54, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s55, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; SI-NEXT: s_or_b32 s56, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; SI-NEXT: s_or_b32 s57, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; SI-NEXT: s_or_b32 s58, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; SI-NEXT: s_or_b32 s59, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s12, 16
|
|
; SI-NEXT: s_or_b32 s60, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s10, 16
|
|
; SI-NEXT: s_or_b32 s61, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s8, 16
|
|
; SI-NEXT: s_or_b32 s62, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s6, 16
|
|
; SI-NEXT: s_or_b32 s63, s4, s5
|
|
; SI-NEXT: s_cbranch_execnz .LBB15_3
|
|
; SI-NEXT: .LBB15_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s17, s17, 3
|
|
; SI-NEXT: s_add_i32 s36, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: s_add_i32 s37, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s19, s19, 3
|
|
; SI-NEXT: s_add_i32 s38, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: s_add_i32 s39, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s21, s21, 3
|
|
; SI-NEXT: s_add_i32 s40, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: s_add_i32 s41, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s42, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s43, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s44, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s34, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s45, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s95, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s46, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s92, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s47, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: s_add_i32 s48, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s80, s80, 3
|
|
; SI-NEXT: s_add_i32 s49, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s80, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s69, s69, 3
|
|
; SI-NEXT: s_add_i32 s50, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s31, s31, 3
|
|
; SI-NEXT: s_add_i32 s51, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s31, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s94, s94, 3
|
|
; SI-NEXT: s_add_i32 s52, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s94, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s91, s91, 3
|
|
; SI-NEXT: s_add_i32 s53, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s88, s88, 3
|
|
; SI-NEXT: s_add_i32 s54, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s77, s77, 3
|
|
; SI-NEXT: s_add_i32 s55, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s75, s75, 3
|
|
; SI-NEXT: s_add_i32 s56, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s73, s73, 3
|
|
; SI-NEXT: s_add_i32 s57, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s15, s15, 3
|
|
; SI-NEXT: s_add_i32 s58, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s13, s13, 3
|
|
; SI-NEXT: s_add_i32 s59, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s12, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s11, s11, 3
|
|
; SI-NEXT: s_add_i32 s60, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s10, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s9, s9, 3
|
|
; SI-NEXT: s_add_i32 s61, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s8, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s7, s7, 3
|
|
; SI-NEXT: s_add_i32 s62, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s6, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s63, s4, 0x30000
|
|
; SI-NEXT: .LBB15_3: ; %end
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_readlane_b32 s87, v28, 31
|
|
; SI-NEXT: v_readlane_b32 s86, v28, 30
|
|
; SI-NEXT: v_readlane_b32 s85, v28, 29
|
|
; SI-NEXT: v_readlane_b32 s84, v28, 28
|
|
; SI-NEXT: v_readlane_b32 s83, v28, 27
|
|
; SI-NEXT: v_readlane_b32 s82, v28, 26
|
|
; SI-NEXT: v_readlane_b32 s81, v28, 25
|
|
; SI-NEXT: v_readlane_b32 s80, v28, 24
|
|
; SI-NEXT: v_readlane_b32 s71, v28, 23
|
|
; SI-NEXT: v_readlane_b32 s70, v28, 22
|
|
; SI-NEXT: v_readlane_b32 s69, v28, 21
|
|
; SI-NEXT: v_readlane_b32 s68, v28, 20
|
|
; SI-NEXT: v_readlane_b32 s67, v28, 19
|
|
; SI-NEXT: v_readlane_b32 s66, v28, 18
|
|
; SI-NEXT: v_readlane_b32 s65, v28, 17
|
|
; SI-NEXT: v_readlane_b32 s64, v28, 16
|
|
; SI-NEXT: v_readlane_b32 s55, v28, 15
|
|
; SI-NEXT: v_readlane_b32 s54, v28, 14
|
|
; SI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB15_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; SI-NEXT: s_branch .LBB15_2
|
|
;
|
|
; VI-LABEL: bitcast_v56i16_to_v28i32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; VI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; VI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; VI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; VI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; VI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; VI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; VI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; VI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; VI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; VI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; VI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; VI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; VI-NEXT: v_writelane_b32 v28, s54, 14
|
|
; VI-NEXT: v_writelane_b32 v28, s55, 15
|
|
; VI-NEXT: v_writelane_b32 v28, s64, 16
|
|
; VI-NEXT: v_writelane_b32 v28, s65, 17
|
|
; VI-NEXT: v_writelane_b32 v28, s66, 18
|
|
; VI-NEXT: v_writelane_b32 v28, s67, 19
|
|
; VI-NEXT: v_writelane_b32 v28, s68, 20
|
|
; VI-NEXT: v_writelane_b32 v28, s69, 21
|
|
; VI-NEXT: v_writelane_b32 v28, s70, 22
|
|
; VI-NEXT: v_writelane_b32 v28, s71, 23
|
|
; VI-NEXT: v_writelane_b32 v28, s80, 24
|
|
; VI-NEXT: v_writelane_b32 v28, s81, 25
|
|
; VI-NEXT: v_writelane_b32 v28, s82, 26
|
|
; VI-NEXT: v_writelane_b32 v28, s83, 27
|
|
; VI-NEXT: v_writelane_b32 v28, s84, 28
|
|
; VI-NEXT: v_writelane_b32 v28, s85, 29
|
|
; VI-NEXT: v_writelane_b32 v28, s86, 30
|
|
; VI-NEXT: v_writelane_b32 v28, s87, 31
|
|
; VI-NEXT: v_readfirstlane_b32 s86, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s87, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s73, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s77, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s88, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s34, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s69, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s81, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s84, v0
|
|
; VI-NEXT: s_lshr_b32 s79, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s67, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s70, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s80, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s82, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s85, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s7, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s11, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s13, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s15, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s10, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s64, s86, 16
|
|
; VI-NEXT: s_lshr_b32 s65, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s66, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s14, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s87, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s73, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s77, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s88, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s91, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s34, 16
|
|
; VI-NEXT: s_lshr_b32 s68, s69, 16
|
|
; VI-NEXT: s_lshr_b32 s71, s81, 16
|
|
; VI-NEXT: s_lshr_b32 s83, s84, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB15_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_or_b32 s36, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_or_b32 s38, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s10, 16
|
|
; VI-NEXT: s_or_b32 s39, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; VI-NEXT: s_or_b32 s37, s40, s41
|
|
; VI-NEXT: s_or_b32 s40, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_or_b32 s41, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s42, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s43, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; VI-NEXT: s_or_b32 s44, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s45, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s5, s67, 16
|
|
; VI-NEXT: s_or_b32 s46, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s47, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s48, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s49, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s84
|
|
; VI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; VI-NEXT: s_or_b32 s50, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s81
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s51, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s69
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s52, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s34
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s53, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s91
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s54, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s88
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s55, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s77
|
|
; VI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; VI-NEXT: s_or_b32 s56, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s73
|
|
; VI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; VI-NEXT: s_or_b32 s57, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; VI-NEXT: s_or_b32 s58, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s87
|
|
; VI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; VI-NEXT: s_or_b32 s59, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; VI-NEXT: s_or_b32 s60, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s5, s66, 16
|
|
; VI-NEXT: s_or_b32 s61, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s5, s65, 16
|
|
; VI-NEXT: s_or_b32 s62, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s86
|
|
; VI-NEXT: s_lshl_b32 s5, s64, 16
|
|
; VI-NEXT: s_or_b32 s63, s4, s5
|
|
; VI-NEXT: s_cbranch_execnz .LBB15_3
|
|
; VI-NEXT: .LBB15_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_and_b32 s5, s17, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s10, s10, 16
|
|
; VI-NEXT: s_or_b32 s5, s10, s5
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s36, s4, 0x30000
|
|
; VI-NEXT: s_add_i32 s37, s5, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s38, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s39, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s40, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s41, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s42, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s43, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s44, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s45, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s67, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s46, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s47, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s48, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s84, s84, 3
|
|
; VI-NEXT: s_add_i32 s49, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s84, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s81, s81, 3
|
|
; VI-NEXT: s_add_i32 s50, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s81, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s69, s69, 3
|
|
; VI-NEXT: s_add_i32 s51, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s34, s34, 3
|
|
; VI-NEXT: s_add_i32 s52, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s34, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s91, s91, 3
|
|
; VI-NEXT: s_add_i32 s53, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s88, s88, 3
|
|
; VI-NEXT: s_add_i32 s54, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s77, s77, 3
|
|
; VI-NEXT: s_add_i32 s55, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s75, s73, 3
|
|
; VI-NEXT: s_add_i32 s56, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s73, s12, 3
|
|
; VI-NEXT: s_add_i32 s57, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s15, s87, 3
|
|
; VI-NEXT: s_add_i32 s58, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s13, s8, 3
|
|
; VI-NEXT: s_add_i32 s59, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s11, s9, 3
|
|
; VI-NEXT: s_add_i32 s60, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s66, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s9, s6, 3
|
|
; VI-NEXT: s_add_i32 s61, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s65, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s7, s86, 3
|
|
; VI-NEXT: s_add_i32 s62, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s64, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s63, s4, 0x30000
|
|
; VI-NEXT: .LBB15_3: ; %end
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_readlane_b32 s87, v28, 31
|
|
; VI-NEXT: v_readlane_b32 s86, v28, 30
|
|
; VI-NEXT: v_readlane_b32 s85, v28, 29
|
|
; VI-NEXT: v_readlane_b32 s84, v28, 28
|
|
; VI-NEXT: v_readlane_b32 s83, v28, 27
|
|
; VI-NEXT: v_readlane_b32 s82, v28, 26
|
|
; VI-NEXT: v_readlane_b32 s81, v28, 25
|
|
; VI-NEXT: v_readlane_b32 s80, v28, 24
|
|
; VI-NEXT: v_readlane_b32 s71, v28, 23
|
|
; VI-NEXT: v_readlane_b32 s70, v28, 22
|
|
; VI-NEXT: v_readlane_b32 s69, v28, 21
|
|
; VI-NEXT: v_readlane_b32 s68, v28, 20
|
|
; VI-NEXT: v_readlane_b32 s67, v28, 19
|
|
; VI-NEXT: v_readlane_b32 s66, v28, 18
|
|
; VI-NEXT: v_readlane_b32 s65, v28, 17
|
|
; VI-NEXT: v_readlane_b32 s64, v28, 16
|
|
; VI-NEXT: v_readlane_b32 s55, v28, 15
|
|
; VI-NEXT: v_readlane_b32 s54, v28, 14
|
|
; VI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; VI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; VI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; VI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; VI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; VI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; VI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; VI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; VI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; VI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; VI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB15_4:
|
|
; VI-NEXT: ; implicit-def: $vgpr29 : SGPR spill to VGPR lane
|
|
; VI-NEXT: s_mov_b32 vcc_lo, s64
|
|
; VI-NEXT: v_writelane_b32 v29, s75, 0
|
|
; VI-NEXT: s_mov_b32 vcc_hi, s65
|
|
; VI-NEXT: s_mov_b32 s75, s73
|
|
; VI-NEXT: s_mov_b32 s73, s12
|
|
; VI-NEXT: v_writelane_b32 v29, s25, 1
|
|
; VI-NEXT: s_mov_b32 s25, s91
|
|
; VI-NEXT: s_mov_b32 s91, s26
|
|
; VI-NEXT: s_mov_b32 s26, s27
|
|
; VI-NEXT: s_mov_b32 s27, s88
|
|
; VI-NEXT: s_mov_b32 s88, s28
|
|
; VI-NEXT: s_mov_b32 s28, s77
|
|
; VI-NEXT: s_mov_b32 s77, s29
|
|
; VI-NEXT: s_mov_b32 s29, s15
|
|
; VI-NEXT: s_mov_b32 s15, s87
|
|
; VI-NEXT: s_mov_b32 s87, s85
|
|
; VI-NEXT: s_mov_b32 s85, s82
|
|
; VI-NEXT: s_mov_b32 s82, s71
|
|
; VI-NEXT: s_mov_b32 s71, s68
|
|
; VI-NEXT: s_mov_b32 s68, s35
|
|
; VI-NEXT: s_mov_b32 s35, s31
|
|
; VI-NEXT: s_mov_b32 s31, s30
|
|
; VI-NEXT: s_mov_b32 s30, s90
|
|
; VI-NEXT: s_mov_b32 s90, s89
|
|
; VI-NEXT: s_mov_b32 s89, s79
|
|
; VI-NEXT: s_mov_b32 s79, s78
|
|
; VI-NEXT: s_mov_b32 s78, s76
|
|
; VI-NEXT: s_mov_b32 s76, s74
|
|
; VI-NEXT: s_mov_b32 s74, s72
|
|
; VI-NEXT: s_mov_b32 s72, s14
|
|
; VI-NEXT: s_mov_b32 s14, s66
|
|
; VI-NEXT: s_mov_b32 s12, s10
|
|
; VI-NEXT: s_mov_b32 s10, s13
|
|
; VI-NEXT: s_mov_b32 s13, s8
|
|
; VI-NEXT: s_mov_b32 s8, s11
|
|
; VI-NEXT: s_mov_b32 s11, s9
|
|
; VI-NEXT: s_mov_b32 s9, s6
|
|
; VI-NEXT: s_mov_b32 s6, s7
|
|
; VI-NEXT: s_mov_b32 s7, s86
|
|
; VI-NEXT: s_mov_b32 s86, s83
|
|
; VI-NEXT: s_mov_b32 s83, s80
|
|
; VI-NEXT: s_mov_b32 s80, s70
|
|
; VI-NEXT: s_mov_b32 s70, s67
|
|
; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; VI-NEXT: s_mov_b32 s67, s70
|
|
; VI-NEXT: s_mov_b32 s70, s80
|
|
; VI-NEXT: s_mov_b32 s80, s83
|
|
; VI-NEXT: s_mov_b32 s83, s86
|
|
; VI-NEXT: s_mov_b32 s86, s7
|
|
; VI-NEXT: s_mov_b32 s7, s6
|
|
; VI-NEXT: s_mov_b32 s6, s9
|
|
; VI-NEXT: s_mov_b32 s9, s11
|
|
; VI-NEXT: s_mov_b32 s11, s8
|
|
; VI-NEXT: s_mov_b32 s8, s13
|
|
; VI-NEXT: s_mov_b32 s13, s10
|
|
; VI-NEXT: s_mov_b32 s10, s12
|
|
; VI-NEXT: s_mov_b32 s66, s14
|
|
; VI-NEXT: s_mov_b32 s14, s72
|
|
; VI-NEXT: s_mov_b32 s72, s74
|
|
; VI-NEXT: s_mov_b32 s74, s76
|
|
; VI-NEXT: s_mov_b32 s76, s78
|
|
; VI-NEXT: s_mov_b32 s78, s79
|
|
; VI-NEXT: s_mov_b32 s79, s89
|
|
; VI-NEXT: s_mov_b32 s89, s90
|
|
; VI-NEXT: s_mov_b32 s90, s30
|
|
; VI-NEXT: s_mov_b32 s30, s31
|
|
; VI-NEXT: s_mov_b32 s31, s35
|
|
; VI-NEXT: s_mov_b32 s35, s68
|
|
; VI-NEXT: s_mov_b32 s68, s71
|
|
; VI-NEXT: s_mov_b32 s71, s82
|
|
; VI-NEXT: s_mov_b32 s82, s85
|
|
; VI-NEXT: s_mov_b32 s85, s87
|
|
; VI-NEXT: s_mov_b32 s87, s15
|
|
; VI-NEXT: s_mov_b32 s15, s29
|
|
; VI-NEXT: s_mov_b32 s29, s77
|
|
; VI-NEXT: s_mov_b32 s77, s28
|
|
; VI-NEXT: s_mov_b32 s28, s88
|
|
; VI-NEXT: s_mov_b32 s88, s27
|
|
; VI-NEXT: s_mov_b32 s27, s26
|
|
; VI-NEXT: s_mov_b32 s26, s91
|
|
; VI-NEXT: s_mov_b32 s91, s25
|
|
; VI-NEXT: v_readlane_b32 s25, v29, 1
|
|
; VI-NEXT: s_mov_b32 s12, s73
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: v_readlane_b32 s75, v29, 0
|
|
; VI-NEXT: s_mov_b32 s65, vcc_hi
|
|
; VI-NEXT: s_mov_b32 s64, vcc_lo
|
|
; VI-NEXT: s_branch .LBB15_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v56i16_to_v28i32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: s_lshr_b32 s12, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s13, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s14, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s15, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s88, v5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v3
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v2
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: s_lshr_b32 s4, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s5, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s6, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s7, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s8, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s9, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s10, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s63, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s62, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s61, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s60, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s59, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s58, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s57, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s56, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s88, 16
|
|
; GFX9-NEXT: s_lshr_b32 s17, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s18, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s19, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s20, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s21, s12, 16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_cmp_lg_u32 s42, 0
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s50, s12, s21
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s51, s13, s20
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s52, s14, s19
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s53, s15, s18
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s54, s16, s17
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s55, s88, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s56, s56, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s57, s57, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s58, s58, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s59, s59, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s60, s60, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s61, s61, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s62, s62, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s63, s63, s72
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB15_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB15_4
|
|
; GFX9-NEXT: .LBB15_2: ; %cmp.true
|
|
; GFX9-NEXT: v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_branch .LBB15_5
|
|
; GFX9-NEXT: .LBB15_3:
|
|
; GFX9-NEXT: s_branch .LBB15_2
|
|
; GFX9-NEXT: .LBB15_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB15_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v56i16_to_v28i32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s44, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s74, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s76, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s77, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s78, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s88, v10
|
|
; GFX11-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s0, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s44, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s56, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s58, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s59, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s60, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s62, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s74, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s76, 16
|
|
; GFX11-NEXT: s_lshr_b32 s90, s77, 16
|
|
; GFX11-NEXT: s_lshr_b32 s91, s78, 16
|
|
; GFX11-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s88, 0
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s78, s91
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s77, s90
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s76, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s74, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s62, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s60, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s59, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s58, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s56, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s44, s57
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB15_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB15_4
|
|
; GFX11-NEXT: .LBB15_2: ; %cmp.true
|
|
; GFX11-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v18, s18, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v19, s19, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v20, s20, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v21, s21, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v22, s22, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v23, s23, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v24, s24, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v25, s25, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v26, s26, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v27, s27, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB15_3:
|
|
; GFX11-NEXT: s_branch .LBB15_2
|
|
; GFX11-NEXT: .LBB15_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <56 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <56 x i16> %a1 to <28 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x i16> %a to <28 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x i32> %phi
|
|
}
|
|
|
|
define <56 x half> @bitcast_v28i32_to_v56f16(<28 x i32> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v28i32_to_v56f16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB16_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB16_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB16_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 3, v25
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 3, v27
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB16_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v47
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v40
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v48
|
|
; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v41
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v53
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v51
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v48
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v38
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28i32_to_v56f16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB16_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB16_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB16_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_add_u32_e32 v27, vcc, 3, v27
|
|
; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
|
|
; VI-NEXT: v_add_u32_e32 v25, vcc, 3, v25
|
|
; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
|
|
; VI-NEXT: v_add_u32_e32 v23, vcc, 3, v23
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
|
|
; VI-NEXT: v_add_u32_e32 v21, vcc, 3, v21
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
|
|
; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB16_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28i32_to_v56f16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB16_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB16_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB16_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: v_add_u32_e32 v27, 3, v27
|
|
; GFX9-NEXT: v_add_u32_e32 v26, 3, v26
|
|
; GFX9-NEXT: v_add_u32_e32 v25, 3, v25
|
|
; GFX9-NEXT: v_add_u32_e32 v24, 3, v24
|
|
; GFX9-NEXT: v_add_u32_e32 v23, 3, v23
|
|
; GFX9-NEXT: v_add_u32_e32 v22, 3, v22
|
|
; GFX9-NEXT: v_add_u32_e32 v21, 3, v21
|
|
; GFX9-NEXT: v_add_u32_e32 v20, 3, v20
|
|
; GFX9-NEXT: v_add_u32_e32 v19, 3, v19
|
|
; GFX9-NEXT: v_add_u32_e32 v18, 3, v18
|
|
; GFX9-NEXT: v_add_u32_e32 v17, 3, v17
|
|
; GFX9-NEXT: v_add_u32_e32 v16, 3, v16
|
|
; GFX9-NEXT: v_add_u32_e32 v15, 3, v15
|
|
; GFX9-NEXT: v_add_u32_e32 v14, 3, v14
|
|
; GFX9-NEXT: v_add_u32_e32 v13, 3, v13
|
|
; GFX9-NEXT: v_add_u32_e32 v12, 3, v12
|
|
; GFX9-NEXT: v_add_u32_e32 v11, 3, v11
|
|
; GFX9-NEXT: v_add_u32_e32 v10, 3, v10
|
|
; GFX9-NEXT: v_add_u32_e32 v9, 3, v9
|
|
; GFX9-NEXT: v_add_u32_e32 v8, 3, v8
|
|
; GFX9-NEXT: v_add_u32_e32 v7, 3, v7
|
|
; GFX9-NEXT: v_add_u32_e32 v6, 3, v6
|
|
; GFX9-NEXT: v_add_u32_e32 v5, 3, v5
|
|
; GFX9-NEXT: v_add_u32_e32 v4, 3, v4
|
|
; GFX9-NEXT: v_add_u32_e32 v3, 3, v3
|
|
; GFX9-NEXT: v_add_u32_e32 v2, 3, v2
|
|
; GFX9-NEXT: v_add_u32_e32 v1, 3, v1
|
|
; GFX9-NEXT: v_add_u32_e32 v0, 3, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB16_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v47, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v46, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v45, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v44, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v43, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v42, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v41, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v40, v7, s4
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v8, v55, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v54, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v53, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v52, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v51, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v49, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v48, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v39, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v38, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v37, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v36, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v35, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v34, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v33, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v32, v23, s4
|
|
; GFX9-NEXT: v_perm_b32 v24, v31, v24, s4
|
|
; GFX9-NEXT: v_perm_b32 v25, v30, v25, s4
|
|
; GFX9-NEXT: v_perm_b32 v26, v29, v26, s4
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s4
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v28i32_to_v56f16:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB16_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v27, 3, v27
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v26, 3, v26
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v25, 3, v25
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v24, 3, v24
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v23, 3, v23
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v22, 3, v22
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v21, 3, v21
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v20, 3, v20
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v19, 3, v19
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v18, 3, v18
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v17, 3, v17
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v16, 3, v16
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v15, 3, v15
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v14, 3, v14
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v13, 3, v13
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v12, 3, v12
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v11, 3, v11
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v10, 3, v10
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, 3, v9
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, 3, v8
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v7, 3, v7
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 3, v6
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v5, 3, v5
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 3, v4
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, 3, v3
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, 3, v2
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v1, 3, v1
|
|
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v0, 3, v0
|
|
; GFX11-TRUE16-NEXT: .LBB16_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v28i32_to_v56f16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr67
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB16_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB16_2: ; %Flow
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB16_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v27, 3, v27
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v26, 3, v26
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v25, 3, v25
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v24, 3, v24
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v23, 3, v23
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v22, 3, v22
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v21, 3, v21
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v20, 3, v20
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v19, 3, v19
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v18, 3, v18
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v17, 3, v17
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v16, 3, v16
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v15, 3, v15
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v14, 3, v14
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v13, 3, v13
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v12, 3, v12
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v11, 3, v11
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v10, 3, v10
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, 3, v9
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, 3, v8
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v7, 3, v7
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, 3, v6
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v5, 3, v5
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v4, 3, v4
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v3, 3, v3
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v2, 3, v2
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v1, 3, v1
|
|
; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v0, 3, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB16_4: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v71, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v69, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v68, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v67, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v66, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v65, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v64, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v55, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v54, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v53, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v52, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v51, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v49, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v39, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v37, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v36, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v35, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v34, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v33, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v32, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v31, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v30, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v29, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <28 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <28 x i32> %a1 to <56 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x i32> %a to <56 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x half> %phi
|
|
}
|
|
|
|
define inreg <56 x half> @bitcast_v28i32_to_v56f16_scalar(<28 x i32> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v28i32_to_v56f16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v14
|
|
; SI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; SI-NEXT: v_readfirstlane_b32 s5, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s40, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v0
|
|
; SI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; SI-NEXT: s_cbranch_scc0 .LBB17_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s30, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s53, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB17_3
|
|
; SI-NEXT: .LBB17_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s17, s17, 3
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: s_add_i32 s19, s19, 3
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: s_add_i32 s21, s21, 3
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s41, s41, 3
|
|
; SI-NEXT: s_add_i32 s40, s40, 3
|
|
; SI-NEXT: s_add_i32 s15, s15, 3
|
|
; SI-NEXT: s_add_i32 s14, s14, 3
|
|
; SI-NEXT: s_add_i32 s13, s13, 3
|
|
; SI-NEXT: s_add_i32 s12, s12, 3
|
|
; SI-NEXT: s_add_i32 s11, s11, 3
|
|
; SI-NEXT: s_add_i32 s10, s10, 3
|
|
; SI-NEXT: s_add_i32 s9, s9, 3
|
|
; SI-NEXT: s_add_i32 s8, s8, 3
|
|
; SI-NEXT: s_add_i32 s7, s7, 3
|
|
; SI-NEXT: s_add_i32 s6, s6, 3
|
|
; SI-NEXT: s_add_i32 s5, s5, 3
|
|
; SI-NEXT: s_add_i32 s4, s4, 3
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b32 s30, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s53, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: .LBB17_3: ; %end
|
|
; SI-NEXT: s_lshl_b32 s43, s92, 16
|
|
; SI-NEXT: s_and_b32 s16, s16, 0xffff
|
|
; SI-NEXT: s_or_b32 s16, s16, s43
|
|
; SI-NEXT: s_and_b32 s17, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s53, 16
|
|
; SI-NEXT: s_or_b32 s17, s17, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s90, 16
|
|
; SI-NEXT: s_and_b32 s18, s18, 0xffff
|
|
; SI-NEXT: s_or_b32 s18, s18, s43
|
|
; SI-NEXT: s_and_b32 s19, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s52, 16
|
|
; SI-NEXT: s_or_b32 s19, s19, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s88, 16
|
|
; SI-NEXT: s_and_b32 s20, s20, 0xffff
|
|
; SI-NEXT: s_or_b32 s20, s20, s43
|
|
; SI-NEXT: s_and_b32 s21, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s51, 16
|
|
; SI-NEXT: s_or_b32 s21, s21, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s78, 16
|
|
; SI-NEXT: s_and_b32 s22, s22, 0xffff
|
|
; SI-NEXT: s_or_b32 s22, s22, s43
|
|
; SI-NEXT: s_and_b32 s23, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s50, 16
|
|
; SI-NEXT: s_or_b32 s23, s23, s43
|
|
; SI-NEXT: s_and_b32 s24, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s76, 16
|
|
; SI-NEXT: s_or_b32 s24, s24, s43
|
|
; SI-NEXT: s_and_b32 s25, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s49, 16
|
|
; SI-NEXT: s_or_b32 s25, s25, s43
|
|
; SI-NEXT: s_and_b32 s26, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s74, 16
|
|
; SI-NEXT: s_or_b32 s26, s26, s43
|
|
; SI-NEXT: s_and_b32 s27, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s48, 16
|
|
; SI-NEXT: s_or_b32 s27, s27, s43
|
|
; SI-NEXT: s_and_b32 s28, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s72, 16
|
|
; SI-NEXT: s_or_b32 s28, s28, s43
|
|
; SI-NEXT: s_and_b32 s29, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s39, 16
|
|
; SI-NEXT: s_or_b32 s29, s29, s43
|
|
; SI-NEXT: s_and_b32 s40, s40, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s62, 16
|
|
; SI-NEXT: s_or_b32 s40, s40, s43
|
|
; SI-NEXT: s_and_b32 s41, s41, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s38, 16
|
|
; SI-NEXT: s_or_b32 s41, s41, s43
|
|
; SI-NEXT: s_and_b32 s14, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s60, 16
|
|
; SI-NEXT: s_or_b32 s14, s14, s43
|
|
; SI-NEXT: s_and_b32 s15, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s37, 16
|
|
; SI-NEXT: s_or_b32 s15, s15, s43
|
|
; SI-NEXT: s_and_b32 s12, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s58, 16
|
|
; SI-NEXT: s_or_b32 s12, s12, s43
|
|
; SI-NEXT: s_and_b32 s13, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s36, 16
|
|
; SI-NEXT: s_or_b32 s13, s13, s43
|
|
; SI-NEXT: s_and_b32 s10, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s56, 16
|
|
; SI-NEXT: s_or_b32 s10, s10, s43
|
|
; SI-NEXT: s_and_b32 s11, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s35, 16
|
|
; SI-NEXT: s_or_b32 s11, s11, s43
|
|
; SI-NEXT: s_and_b32 s8, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s46, 16
|
|
; SI-NEXT: s_or_b32 s8, s8, s43
|
|
; SI-NEXT: s_and_b32 s9, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s34, 16
|
|
; SI-NEXT: s_or_b32 s9, s9, s43
|
|
; SI-NEXT: s_and_b32 s6, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s44, 16
|
|
; SI-NEXT: s_and_b32 s4, s4, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s42, 16
|
|
; SI-NEXT: s_or_b32 s6, s6, s43
|
|
; SI-NEXT: s_and_b32 s7, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s31, 16
|
|
; SI-NEXT: s_or_b32 s4, s4, s42
|
|
; SI-NEXT: s_and_b32 s5, s5, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s30, 16
|
|
; SI-NEXT: s_or_b32 s7, s7, s43
|
|
; SI-NEXT: s_or_b32 s5, s5, s42
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s40
|
|
; SI-NEXT: v_mov_b32_e32 v15, s41
|
|
; SI-NEXT: v_mov_b32_e32 v16, s14
|
|
; SI-NEXT: v_mov_b32_e32 v17, s15
|
|
; SI-NEXT: v_mov_b32_e32 v18, s12
|
|
; SI-NEXT: v_mov_b32_e32 v19, s13
|
|
; SI-NEXT: v_mov_b32_e32 v20, s10
|
|
; SI-NEXT: v_mov_b32_e32 v21, s11
|
|
; SI-NEXT: v_mov_b32_e32 v22, s8
|
|
; SI-NEXT: v_mov_b32_e32 v23, s9
|
|
; SI-NEXT: v_mov_b32_e32 v24, s6
|
|
; SI-NEXT: v_mov_b32_e32 v25, s7
|
|
; SI-NEXT: v_mov_b32_e32 v26, s4
|
|
; SI-NEXT: v_mov_b32_e32 v27, s5
|
|
; SI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB17_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr92
|
|
; SI-NEXT: ; implicit-def: $sgpr53
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: ; implicit-def: $sgpr52
|
|
; SI-NEXT: ; implicit-def: $sgpr88
|
|
; SI-NEXT: ; implicit-def: $sgpr51
|
|
; SI-NEXT: ; implicit-def: $sgpr78
|
|
; SI-NEXT: ; implicit-def: $sgpr50
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr49
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr48
|
|
; SI-NEXT: ; implicit-def: $sgpr72
|
|
; SI-NEXT: ; implicit-def: $sgpr39
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr38
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr37
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr36
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr35
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr31
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: s_branch .LBB17_2
|
|
;
|
|
; VI-LABEL: bitcast_v28i32_to_v56f16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB17_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s44, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s42, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB17_3
|
|
; VI-NEXT: .LBB17_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s6, s6, 3
|
|
; VI-NEXT: s_add_i32 s7, s7, 3
|
|
; VI-NEXT: s_add_i32 s8, s8, 3
|
|
; VI-NEXT: s_add_i32 s9, s9, 3
|
|
; VI-NEXT: s_add_i32 s10, s10, 3
|
|
; VI-NEXT: s_add_i32 s11, s11, 3
|
|
; VI-NEXT: s_add_i32 s12, s12, 3
|
|
; VI-NEXT: s_add_i32 s13, s13, 3
|
|
; VI-NEXT: s_add_i32 s14, s14, 3
|
|
; VI-NEXT: s_add_i32 s15, s15, 3
|
|
; VI-NEXT: s_add_i32 s40, s40, 3
|
|
; VI-NEXT: s_add_i32 s41, s41, 3
|
|
; VI-NEXT: s_add_i32 s42, s42, 3
|
|
; VI-NEXT: s_add_i32 s43, s43, 3
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: s_lshr_b32 s44, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s42, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: .LBB17_3: ; %end
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s16, s34, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s16
|
|
; VI-NEXT: s_and_b32 s16, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s17, s31, 16
|
|
; VI-NEXT: s_or_b32 s16, s16, s17
|
|
; VI-NEXT: s_and_b32 s17, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s18, s30, 16
|
|
; VI-NEXT: s_or_b32 s17, s17, s18
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s19, s91, 16
|
|
; VI-NEXT: s_or_b32 s18, s18, s19
|
|
; VI-NEXT: s_and_b32 s19, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s20, s90, 16
|
|
; VI-NEXT: s_or_b32 s19, s19, s20
|
|
; VI-NEXT: s_and_b32 s20, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s21, s89, 16
|
|
; VI-NEXT: s_or_b32 s20, s20, s21
|
|
; VI-NEXT: s_and_b32 s21, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s22, s88, 16
|
|
; VI-NEXT: s_or_b32 s21, s21, s22
|
|
; VI-NEXT: s_and_b32 s22, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s23, s79, 16
|
|
; VI-NEXT: s_or_b32 s22, s22, s23
|
|
; VI-NEXT: s_and_b32 s23, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s24, s78, 16
|
|
; VI-NEXT: s_or_b32 s23, s23, s24
|
|
; VI-NEXT: s_and_b32 s24, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s25, s77, 16
|
|
; VI-NEXT: s_or_b32 s24, s24, s25
|
|
; VI-NEXT: s_and_b32 s25, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s26, s76, 16
|
|
; VI-NEXT: s_or_b32 s25, s25, s26
|
|
; VI-NEXT: s_and_b32 s26, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s27, s75, 16
|
|
; VI-NEXT: s_or_b32 s26, s26, s27
|
|
; VI-NEXT: s_and_b32 s27, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s28, s74, 16
|
|
; VI-NEXT: s_or_b32 s27, s27, s28
|
|
; VI-NEXT: s_and_b32 s28, 0xffff, s43
|
|
; VI-NEXT: s_lshl_b32 s29, s73, 16
|
|
; VI-NEXT: s_or_b32 s28, s28, s29
|
|
; VI-NEXT: s_and_b32 s29, 0xffff, s42
|
|
; VI-NEXT: s_lshl_b32 s42, s72, 16
|
|
; VI-NEXT: s_or_b32 s29, s29, s42
|
|
; VI-NEXT: s_and_b32 s41, 0xffff, s41
|
|
; VI-NEXT: s_lshl_b32 s42, s63, 16
|
|
; VI-NEXT: s_or_b32 s41, s41, s42
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s40
|
|
; VI-NEXT: s_lshl_b32 s42, s62, 16
|
|
; VI-NEXT: s_or_b32 s40, s40, s42
|
|
; VI-NEXT: s_and_b32 s15, 0xffff, s15
|
|
; VI-NEXT: s_lshl_b32 s42, s61, 16
|
|
; VI-NEXT: s_or_b32 s15, s15, s42
|
|
; VI-NEXT: s_and_b32 s14, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s42, s60, 16
|
|
; VI-NEXT: s_or_b32 s14, s14, s42
|
|
; VI-NEXT: s_and_b32 s13, 0xffff, s13
|
|
; VI-NEXT: s_lshl_b32 s42, s59, 16
|
|
; VI-NEXT: s_or_b32 s13, s13, s42
|
|
; VI-NEXT: s_and_b32 s12, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s42, s58, 16
|
|
; VI-NEXT: s_or_b32 s12, s12, s42
|
|
; VI-NEXT: s_and_b32 s11, 0xffff, s11
|
|
; VI-NEXT: s_lshl_b32 s42, s57, 16
|
|
; VI-NEXT: s_or_b32 s11, s11, s42
|
|
; VI-NEXT: s_and_b32 s10, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s42, s56, 16
|
|
; VI-NEXT: s_or_b32 s10, s10, s42
|
|
; VI-NEXT: s_and_b32 s9, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s42, s47, 16
|
|
; VI-NEXT: s_or_b32 s9, s9, s42
|
|
; VI-NEXT: s_and_b32 s8, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s42, s46, 16
|
|
; VI-NEXT: s_or_b32 s8, s8, s42
|
|
; VI-NEXT: s_and_b32 s7, 0xffff, s7
|
|
; VI-NEXT: s_lshl_b32 s42, s45, 16
|
|
; VI-NEXT: s_or_b32 s7, s7, s42
|
|
; VI-NEXT: s_and_b32 s6, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s42, s44, 16
|
|
; VI-NEXT: s_or_b32 s6, s6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s16
|
|
; VI-NEXT: v_mov_b32_e32 v3, s17
|
|
; VI-NEXT: v_mov_b32_e32 v4, s18
|
|
; VI-NEXT: v_mov_b32_e32 v5, s19
|
|
; VI-NEXT: v_mov_b32_e32 v6, s20
|
|
; VI-NEXT: v_mov_b32_e32 v7, s21
|
|
; VI-NEXT: v_mov_b32_e32 v8, s22
|
|
; VI-NEXT: v_mov_b32_e32 v9, s23
|
|
; VI-NEXT: v_mov_b32_e32 v10, s24
|
|
; VI-NEXT: v_mov_b32_e32 v11, s25
|
|
; VI-NEXT: v_mov_b32_e32 v12, s26
|
|
; VI-NEXT: v_mov_b32_e32 v13, s27
|
|
; VI-NEXT: v_mov_b32_e32 v14, s28
|
|
; VI-NEXT: v_mov_b32_e32 v15, s29
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB17_4:
|
|
; VI-NEXT: ; implicit-def: $sgpr35
|
|
; VI-NEXT: ; implicit-def: $sgpr34
|
|
; VI-NEXT: ; implicit-def: $sgpr31
|
|
; VI-NEXT: ; implicit-def: $sgpr30
|
|
; VI-NEXT: ; implicit-def: $sgpr91
|
|
; VI-NEXT: ; implicit-def: $sgpr90
|
|
; VI-NEXT: ; implicit-def: $sgpr89
|
|
; VI-NEXT: ; implicit-def: $sgpr88
|
|
; VI-NEXT: ; implicit-def: $sgpr79
|
|
; VI-NEXT: ; implicit-def: $sgpr78
|
|
; VI-NEXT: ; implicit-def: $sgpr77
|
|
; VI-NEXT: ; implicit-def: $sgpr76
|
|
; VI-NEXT: ; implicit-def: $sgpr75
|
|
; VI-NEXT: ; implicit-def: $sgpr74
|
|
; VI-NEXT: ; implicit-def: $sgpr73
|
|
; VI-NEXT: ; implicit-def: $sgpr72
|
|
; VI-NEXT: ; implicit-def: $sgpr63
|
|
; VI-NEXT: ; implicit-def: $sgpr62
|
|
; VI-NEXT: ; implicit-def: $sgpr61
|
|
; VI-NEXT: ; implicit-def: $sgpr60
|
|
; VI-NEXT: ; implicit-def: $sgpr59
|
|
; VI-NEXT: ; implicit-def: $sgpr58
|
|
; VI-NEXT: ; implicit-def: $sgpr57
|
|
; VI-NEXT: ; implicit-def: $sgpr56
|
|
; VI-NEXT: ; implicit-def: $sgpr47
|
|
; VI-NEXT: ; implicit-def: $sgpr46
|
|
; VI-NEXT: ; implicit-def: $sgpr45
|
|
; VI-NEXT: ; implicit-def: $sgpr44
|
|
; VI-NEXT: s_branch .LBB17_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v28i32_to_v56f16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB17_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s44, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s42, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s43, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB17_3
|
|
; GFX9-NEXT: .LBB17_2: ; %cmp.true
|
|
; GFX9-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX9-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX9-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX9-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX9-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX9-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX9-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX9-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX9-NEXT: s_add_i32 s14, s14, 3
|
|
; GFX9-NEXT: s_add_i32 s15, s15, 3
|
|
; GFX9-NEXT: s_add_i32 s40, s40, 3
|
|
; GFX9-NEXT: s_add_i32 s41, s41, 3
|
|
; GFX9-NEXT: s_add_i32 s42, s42, 3
|
|
; GFX9-NEXT: s_add_i32 s43, s43, 3
|
|
; GFX9-NEXT: s_add_i32 s29, s29, 3
|
|
; GFX9-NEXT: s_add_i32 s28, s28, 3
|
|
; GFX9-NEXT: s_add_i32 s27, s27, 3
|
|
; GFX9-NEXT: s_add_i32 s26, s26, 3
|
|
; GFX9-NEXT: s_add_i32 s25, s25, 3
|
|
; GFX9-NEXT: s_add_i32 s24, s24, 3
|
|
; GFX9-NEXT: s_add_i32 s23, s23, 3
|
|
; GFX9-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX9-NEXT: s_add_i32 s21, s21, 3
|
|
; GFX9-NEXT: s_add_i32 s20, s20, 3
|
|
; GFX9-NEXT: s_add_i32 s19, s19, 3
|
|
; GFX9-NEXT: s_add_i32 s18, s18, 3
|
|
; GFX9-NEXT: s_add_i32 s17, s17, 3
|
|
; GFX9-NEXT: s_add_i32 s16, s16, 3
|
|
; GFX9-NEXT: s_lshr_b32 s44, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s42, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s43, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: .LBB17_3: ; %end
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s95
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s94
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s93
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s92
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s91
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s90
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s88
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s28, s43, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s29, s42, s72
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s41, s63
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s40, s62
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s15, s61
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s14, s60
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s13, s59
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s12, s58
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s57
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s56
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s47
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s46
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s45
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s4
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s5
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB17_4:
|
|
; GFX9-NEXT: ; implicit-def: $sgpr95
|
|
; GFX9-NEXT: ; implicit-def: $sgpr94
|
|
; GFX9-NEXT: ; implicit-def: $sgpr93
|
|
; GFX9-NEXT: ; implicit-def: $sgpr92
|
|
; GFX9-NEXT: ; implicit-def: $sgpr91
|
|
; GFX9-NEXT: ; implicit-def: $sgpr90
|
|
; GFX9-NEXT: ; implicit-def: $sgpr89
|
|
; GFX9-NEXT: ; implicit-def: $sgpr88
|
|
; GFX9-NEXT: ; implicit-def: $sgpr79
|
|
; GFX9-NEXT: ; implicit-def: $sgpr78
|
|
; GFX9-NEXT: ; implicit-def: $sgpr77
|
|
; GFX9-NEXT: ; implicit-def: $sgpr76
|
|
; GFX9-NEXT: ; implicit-def: $sgpr75
|
|
; GFX9-NEXT: ; implicit-def: $sgpr74
|
|
; GFX9-NEXT: ; implicit-def: $sgpr73
|
|
; GFX9-NEXT: ; implicit-def: $sgpr72
|
|
; GFX9-NEXT: ; implicit-def: $sgpr63
|
|
; GFX9-NEXT: ; implicit-def: $sgpr62
|
|
; GFX9-NEXT: ; implicit-def: $sgpr61
|
|
; GFX9-NEXT: ; implicit-def: $sgpr60
|
|
; GFX9-NEXT: ; implicit-def: $sgpr59
|
|
; GFX9-NEXT: ; implicit-def: $sgpr58
|
|
; GFX9-NEXT: ; implicit-def: $sgpr57
|
|
; GFX9-NEXT: ; implicit-def: $sgpr56
|
|
; GFX9-NEXT: ; implicit-def: $sgpr47
|
|
; GFX9-NEXT: ; implicit-def: $sgpr46
|
|
; GFX9-NEXT: ; implicit-def: $sgpr45
|
|
; GFX9-NEXT: ; implicit-def: $sgpr44
|
|
; GFX9-NEXT: s_branch .LBB17_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v28i32_to_v56f16_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-NEXT: s_mov_b32 s90, 0
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB17_4
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-NEXT: s_lshr_b32 s14, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s74, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s76, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s77, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s78, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s88, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s0, 16
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s90
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB17_3
|
|
; GFX11-NEXT: .LBB17_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_i32 s4, s4, 3
|
|
; GFX11-NEXT: s_add_i32 s5, s5, 3
|
|
; GFX11-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX11-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX11-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX11-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX11-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX11-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX11-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX11-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX11-NEXT: s_add_i32 s29, s29, 3
|
|
; GFX11-NEXT: s_add_i32 s28, s28, 3
|
|
; GFX11-NEXT: s_add_i32 s27, s27, 3
|
|
; GFX11-NEXT: s_add_i32 s26, s26, 3
|
|
; GFX11-NEXT: s_add_i32 s25, s25, 3
|
|
; GFX11-NEXT: s_add_i32 s24, s24, 3
|
|
; GFX11-NEXT: s_add_i32 s23, s23, 3
|
|
; GFX11-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX11-NEXT: s_add_i32 s21, s21, 3
|
|
; GFX11-NEXT: s_add_i32 s20, s20, 3
|
|
; GFX11-NEXT: s_add_i32 s19, s19, 3
|
|
; GFX11-NEXT: s_add_i32 s18, s18, 3
|
|
; GFX11-NEXT: s_add_i32 s17, s17, 3
|
|
; GFX11-NEXT: s_add_i32 s16, s16, 3
|
|
; GFX11-NEXT: s_add_i32 s3, s3, 3
|
|
; GFX11-NEXT: s_add_i32 s2, s2, 3
|
|
; GFX11-NEXT: s_add_i32 s1, s1, 3
|
|
; GFX11-NEXT: s_add_i32 s0, s0, 3
|
|
; GFX11-NEXT: s_lshr_b32 s14, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s74, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s76, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s77, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s78, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s88, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s0, 16
|
|
; GFX11-NEXT: .LBB17_3: ; %end
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s88
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s78
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s77
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s76
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s74
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s62
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s60
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s59
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s58
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s57
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s56
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s44
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s40
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s14
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB17_4:
|
|
; GFX11-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-NEXT: ; implicit-def: $sgpr14
|
|
; GFX11-NEXT: s_branch .LBB17_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <28 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <28 x i32> %a1 to <56 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x i32> %a to <56 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x half> %phi
|
|
}
|
|
|
|
define <28 x i32> @bitcast_v56f16_to_v28i32(<56 x half> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v56f16_to_v28i32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_mov_b32_e32 v59, v0
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v22
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v21
|
|
; SI-NEXT: v_mov_b32_e32 v48, v19
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v20
|
|
; SI-NEXT: v_mov_b32_e32 v49, v18
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; SI-NEXT: v_mov_b32_e32 v50, v17
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; SI-NEXT: v_mov_b32_e32 v51, v16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; SI-NEXT: v_mov_b32_e32 v52, v15
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; SI-NEXT: v_mov_b32_e32 v53, v14
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; SI-NEXT: v_mov_b32_e32 v54, v13
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; SI-NEXT: v_mov_b32_e32 v55, v12
|
|
; SI-NEXT: v_mov_b32_e32 v40, v11
|
|
; SI-NEXT: v_mov_b32_e32 v41, v10
|
|
; SI-NEXT: v_mov_b32_e32 v42, v9
|
|
; SI-NEXT: v_mov_b32_e32 v43, v8
|
|
; SI-NEXT: v_mov_b32_e32 v44, v7
|
|
; SI-NEXT: v_mov_b32_e32 v45, v6
|
|
; SI-NEXT: v_mov_b32_e32 v46, v5
|
|
; SI-NEXT: v_mov_b32_e32 v47, v4
|
|
; SI-NEXT: v_mov_b32_e32 v56, v3
|
|
; SI-NEXT: v_mov_b32_e32 v57, v2
|
|
; SI-NEXT: v_mov_b32_e32 v58, v1
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v40
|
|
; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v41
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v42
|
|
; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v43
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v44
|
|
; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v46
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v47
|
|
; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v56
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v57
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v58
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v59
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB18_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v59
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v39
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v58
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v3
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v57
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v62
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v47
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v46
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v45
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v61
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v43
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v60
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v41
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v63
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v11
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v40
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v55
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v53
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v52
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v51
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v50
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v49
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v27
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: .LBB18_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB18_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v32
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v58
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v56
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v47
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v62
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v46
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v45
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v37
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v44
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v43
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v36
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v42
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v41
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v61
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v40
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v55
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; SI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v35
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v54
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v53
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v60
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v52
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v51
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v34
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v49
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v48
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v63
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v33
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v29, 0x38000000, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v16
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v18, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v19
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v21, v19
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v27, v25
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_or_b32_e32 v24, v26, v24
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v27, v29, v27
|
|
; SI-NEXT: .LBB18_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56f16_to_v28i32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_mov_b32_e32 v32, v27
|
|
; VI-NEXT: v_mov_b32_e32 v33, v26
|
|
; VI-NEXT: v_mov_b32_e32 v34, v25
|
|
; VI-NEXT: v_mov_b32_e32 v35, v24
|
|
; VI-NEXT: v_mov_b32_e32 v36, v23
|
|
; VI-NEXT: v_mov_b32_e32 v37, v22
|
|
; VI-NEXT: v_mov_b32_e32 v38, v21
|
|
; VI-NEXT: v_mov_b32_e32 v39, v20
|
|
; VI-NEXT: v_mov_b32_e32 v48, v19
|
|
; VI-NEXT: v_mov_b32_e32 v49, v18
|
|
; VI-NEXT: v_mov_b32_e32 v50, v17
|
|
; VI-NEXT: v_mov_b32_e32 v51, v16
|
|
; VI-NEXT: v_mov_b32_e32 v52, v15
|
|
; VI-NEXT: v_mov_b32_e32 v53, v14
|
|
; VI-NEXT: v_mov_b32_e32 v54, v13
|
|
; VI-NEXT: v_mov_b32_e32 v55, v12
|
|
; VI-NEXT: v_mov_b32_e32 v40, v11
|
|
; VI-NEXT: v_mov_b32_e32 v41, v10
|
|
; VI-NEXT: v_mov_b32_e32 v42, v9
|
|
; VI-NEXT: v_mov_b32_e32 v43, v8
|
|
; VI-NEXT: v_mov_b32_e32 v44, v7
|
|
; VI-NEXT: v_mov_b32_e32 v45, v6
|
|
; VI-NEXT: v_mov_b32_e32 v46, v5
|
|
; VI-NEXT: v_mov_b32_e32 v47, v4
|
|
; VI-NEXT: v_mov_b32_e32 v56, v3
|
|
; VI-NEXT: v_mov_b32_e32 v57, v2
|
|
; VI-NEXT: v_mov_b32_e32 v58, v1
|
|
; VI-NEXT: v_mov_b32_e32 v59, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB18_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_mov_b32_e32 v27, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v27, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v27, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v27, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v27, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v27, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v27, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v27, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v27, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v27, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v27, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v27, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v27, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v27, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v27, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v27, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v27, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v27, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v27, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v27, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v27, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v27, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v24, v27, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v25, v27, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v26, v27, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v27, v27, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v56, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v47, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v46, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v45, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v44, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v43, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v42, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v41, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v40, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v53, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v52, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v51, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v50, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v49, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v39, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v37, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v36, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v35, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v34, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v33, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v32, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: ; implicit-def: $vgpr59
|
|
; VI-NEXT: ; implicit-def: $vgpr58
|
|
; VI-NEXT: ; implicit-def: $vgpr57
|
|
; VI-NEXT: ; implicit-def: $vgpr56
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: .LBB18_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB18_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; VI-NEXT: v_add_f16_sdwa v0, v59, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v59
|
|
; VI-NEXT: v_add_f16_sdwa v2, v58, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v58
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v2, v57, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v57
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v3, v56, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v56
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_add_f16_sdwa v4, v47, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v47
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_add_f16_sdwa v5, v46, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v46
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_add_f16_sdwa v6, v45, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v45
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_add_f16_sdwa v7, v44, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v44
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_add_f16_sdwa v8, v43, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v43
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_add_f16_sdwa v9, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, 0x200, v42
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v41, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, 0x200, v41
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_add_f16_sdwa v11, v40, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, 0x200, v40
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_add_f16_sdwa v12, v55, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, 0x200, v55
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_add_f16_sdwa v13, v54, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v54
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_add_f16_sdwa v14, v53, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v53
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v52, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v52
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v51, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v51
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v50, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v50
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v49, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v48, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v39, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v38, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v37, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v23, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v36, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: v_add_f16_sdwa v24, v35, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v25, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; VI-NEXT: v_add_f16_sdwa v25, v34, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v26, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; VI-NEXT: v_add_f16_sdwa v26, v33, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v26, v28, v26
|
|
; VI-NEXT: v_add_f16_sdwa v27, v32, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; VI-NEXT: .LBB18_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56f16_to_v28i32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_mov_b32_e32 v59, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v22
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v13
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v12
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v11
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v10
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v40
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v56, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v57, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v58, v1
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v41
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v47
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v56
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v57
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v58
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v59
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB18_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v37
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: .LBB18_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB18_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: s_movk_i32 s7, 0x200
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v0, v0, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, v1, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, v2, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, v3, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, v4, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, v5, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, v12, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, v13, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, v14, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, v15, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, v16, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, v17, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, v18, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, v19, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, v20, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, v21, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, v22, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, v23, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v24, v24, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v25, v25, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v26, v26, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v27, v27, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB18_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v28i32:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB18_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: .LBB18_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56f16_to_v28i32:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v69, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v71, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v80, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v68, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v67, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v66, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v65, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v64, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v54, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v53, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v52, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v51, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v50, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v49, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v48, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v39, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v38, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v37, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v36, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v35, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v34, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v33, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v32, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v31, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v30, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v29, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB18_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: .LBB18_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <56 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <56 x half> %a1 to <28 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x half> %a to <28 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x i32> %phi
|
|
}
|
|
|
|
define inreg <28 x i32> @bitcast_v56f16_to_v28i32_scalar(<56 x half> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v56f16_to_v28i32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v32, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v32, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v32, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v32, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v32, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v32, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v32, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v32, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v32, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v32, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v32, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v32, s51, 11
|
|
; SI-NEXT: v_writelane_b32 v32, s52, 12
|
|
; SI-NEXT: v_writelane_b32 v32, s53, 13
|
|
; SI-NEXT: v_writelane_b32 v32, s54, 14
|
|
; SI-NEXT: v_writelane_b32 v32, s55, 15
|
|
; SI-NEXT: v_writelane_b32 v32, s64, 16
|
|
; SI-NEXT: v_writelane_b32 v32, s65, 17
|
|
; SI-NEXT: v_writelane_b32 v32, s66, 18
|
|
; SI-NEXT: v_writelane_b32 v32, s67, 19
|
|
; SI-NEXT: v_writelane_b32 v32, s68, 20
|
|
; SI-NEXT: v_writelane_b32 v32, s69, 21
|
|
; SI-NEXT: v_writelane_b32 v32, s70, 22
|
|
; SI-NEXT: v_writelane_b32 v32, s71, 23
|
|
; SI-NEXT: v_writelane_b32 v32, s80, 24
|
|
; SI-NEXT: v_writelane_b32 v32, s81, 25
|
|
; SI-NEXT: v_writelane_b32 v32, s82, 26
|
|
; SI-NEXT: v_writelane_b32 v32, s83, 27
|
|
; SI-NEXT: v_writelane_b32 v32, s84, 28
|
|
; SI-NEXT: v_writelane_b32 v32, s85, 29
|
|
; SI-NEXT: v_writelane_b32 v32, s86, 30
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s72, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s74, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s76, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s78, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s88, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s90, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s92, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s95, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s34, v0
|
|
; SI-NEXT: v_writelane_b32 v32, s87, 31
|
|
; SI-NEXT: s_lshr_b32 s94, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s69, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s70, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s71, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s80, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s81, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s82, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s83, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s84, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s85, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s86, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s87, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s7, s6, 16
|
|
; SI-NEXT: s_lshr_b32 s9, s8, 16
|
|
; SI-NEXT: s_lshr_b32 s11, s10, 16
|
|
; SI-NEXT: s_lshr_b32 s13, s12, 16
|
|
; SI-NEXT: s_lshr_b32 s15, s14, 16
|
|
; SI-NEXT: s_lshr_b32 s73, s72, 16
|
|
; SI-NEXT: s_lshr_b32 s75, s74, 16
|
|
; SI-NEXT: s_lshr_b32 s77, s76, 16
|
|
; SI-NEXT: s_lshr_b32 s79, s78, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s88, 16
|
|
; SI-NEXT: s_lshr_b32 s91, s90, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s92, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s95, 16
|
|
; SI-NEXT: s_lshr_b32 s68, s34, 16
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB19_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s36, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s37, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s38, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s39, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s40, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s41, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s42, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; SI-NEXT: s_or_b32 s43, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s44, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s45, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s69, 16
|
|
; SI-NEXT: s_or_b32 s46, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s47, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s48, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s94, 16
|
|
; SI-NEXT: s_or_b32 s49, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s34, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s50, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s95, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; SI-NEXT: s_or_b32 s51, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s92, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s52, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s90, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s91, 16
|
|
; SI-NEXT: s_or_b32 s53, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s54, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s78, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s55, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s76, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s77, 16
|
|
; SI-NEXT: s_or_b32 s56, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s74, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; SI-NEXT: s_or_b32 s57, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s72, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s73, 16
|
|
; SI-NEXT: s_or_b32 s58, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; SI-NEXT: s_or_b32 s59, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; SI-NEXT: s_or_b32 s60, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; SI-NEXT: s_or_b32 s61, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s9, 16
|
|
; SI-NEXT: s_or_b32 s62, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; SI-NEXT: s_or_b32 s63, s4, s5
|
|
; SI-NEXT: s_cbranch_execnz .LBB19_4
|
|
; SI-NEXT: .LBB19_2: ; %cmp.true
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, s87
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s86
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s17
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s85
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s18
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s84
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s19
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s83
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s20
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_or_b32_e32 v3, v5, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s82
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_or_b32_e32 v4, v6, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s81
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s22
|
|
; SI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s80
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s23
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_or_b32_e32 v6, v8, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s71
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v9, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s70
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s25
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s69
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s26
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_or_b32_e32 v9, v11, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v12, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s30
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s28
|
|
; SI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s94
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s29
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s68
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v13, v15, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s31
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s95
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s93
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s92
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s91
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v18, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s90
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s89
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s88
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s79
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s78
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s77
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v21, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s76
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s75
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s74
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s73
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s72
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s13
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s12
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s10
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: v_or_b32_e32 v24, v26, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v27, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, s7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, s6
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v29, 0x38000000, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v27, v29, v27
|
|
; SI-NEXT: s_branch .LBB19_5
|
|
; SI-NEXT: .LBB19_3:
|
|
; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; SI-NEXT: s_branch .LBB19_2
|
|
; SI-NEXT: .LBB19_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_mov_b32_e32 v28, s64
|
|
; SI-NEXT: v_mov_b32_e32 v29, s65
|
|
; SI-NEXT: v_mov_b32_e32 v30, s66
|
|
; SI-NEXT: v_mov_b32_e32 v31, s67
|
|
; SI-NEXT: .LBB19_5: ; %end
|
|
; SI-NEXT: v_readlane_b32 s87, v32, 31
|
|
; SI-NEXT: v_readlane_b32 s86, v32, 30
|
|
; SI-NEXT: v_readlane_b32 s85, v32, 29
|
|
; SI-NEXT: v_readlane_b32 s84, v32, 28
|
|
; SI-NEXT: v_readlane_b32 s83, v32, 27
|
|
; SI-NEXT: v_readlane_b32 s82, v32, 26
|
|
; SI-NEXT: v_readlane_b32 s81, v32, 25
|
|
; SI-NEXT: v_readlane_b32 s80, v32, 24
|
|
; SI-NEXT: v_readlane_b32 s71, v32, 23
|
|
; SI-NEXT: v_readlane_b32 s70, v32, 22
|
|
; SI-NEXT: v_readlane_b32 s69, v32, 21
|
|
; SI-NEXT: v_readlane_b32 s68, v32, 20
|
|
; SI-NEXT: v_readlane_b32 s67, v32, 19
|
|
; SI-NEXT: v_readlane_b32 s66, v32, 18
|
|
; SI-NEXT: v_readlane_b32 s65, v32, 17
|
|
; SI-NEXT: v_readlane_b32 s64, v32, 16
|
|
; SI-NEXT: v_readlane_b32 s55, v32, 15
|
|
; SI-NEXT: v_readlane_b32 s54, v32, 14
|
|
; SI-NEXT: v_readlane_b32 s53, v32, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v32, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v32, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v32, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v32, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v32, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v32, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v32, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v32, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v32, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v32, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v32, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v32, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v32, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56f16_to_v28i32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v32, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s31, 1
|
|
; VI-NEXT: v_writelane_b32 v32, s34, 2
|
|
; VI-NEXT: v_writelane_b32 v32, s35, 3
|
|
; VI-NEXT: v_writelane_b32 v32, s36, 4
|
|
; VI-NEXT: v_writelane_b32 v32, s37, 5
|
|
; VI-NEXT: v_writelane_b32 v32, s38, 6
|
|
; VI-NEXT: v_writelane_b32 v32, s39, 7
|
|
; VI-NEXT: v_writelane_b32 v32, s48, 8
|
|
; VI-NEXT: v_writelane_b32 v32, s49, 9
|
|
; VI-NEXT: v_writelane_b32 v32, s50, 10
|
|
; VI-NEXT: v_writelane_b32 v32, s51, 11
|
|
; VI-NEXT: v_writelane_b32 v32, s52, 12
|
|
; VI-NEXT: v_writelane_b32 v32, s53, 13
|
|
; VI-NEXT: v_writelane_b32 v32, s54, 14
|
|
; VI-NEXT: v_writelane_b32 v32, s55, 15
|
|
; VI-NEXT: v_writelane_b32 v32, s64, 16
|
|
; VI-NEXT: v_writelane_b32 v32, s65, 17
|
|
; VI-NEXT: v_writelane_b32 v32, s66, 18
|
|
; VI-NEXT: v_writelane_b32 v32, s67, 19
|
|
; VI-NEXT: v_writelane_b32 v32, s68, 20
|
|
; VI-NEXT: v_writelane_b32 v32, s69, 21
|
|
; VI-NEXT: v_writelane_b32 v32, s70, 22
|
|
; VI-NEXT: v_writelane_b32 v32, s71, 23
|
|
; VI-NEXT: v_writelane_b32 v32, s80, 24
|
|
; VI-NEXT: v_writelane_b32 v32, s81, 25
|
|
; VI-NEXT: v_writelane_b32 v32, s82, 26
|
|
; VI-NEXT: v_writelane_b32 v32, s83, 27
|
|
; VI-NEXT: v_writelane_b32 v32, s84, 28
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; VI-NEXT: v_writelane_b32 v32, s85, 29
|
|
; VI-NEXT: s_lshr_b32 s15, s8, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v11
|
|
; VI-NEXT: ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
|
|
; VI-NEXT: v_writelane_b32 v32, s86, 30
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: s_lshr_b32 s61, s10, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s72, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s74, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s76, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s79, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s34, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s69, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s80, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s83, v0
|
|
; VI-NEXT: v_writelane_b32 v33, s15, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s87, 31
|
|
; VI-NEXT: s_lshr_b32 s56, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s68, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s70, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s81, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s84, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s86, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s87, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s7, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s9, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s11, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s13, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s72, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s74, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s76, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s79, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s91, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s34, 16
|
|
; VI-NEXT: s_lshr_b32 s71, s69, 16
|
|
; VI-NEXT: s_lshr_b32 s82, s80, 16
|
|
; VI-NEXT: s_lshr_b32 s85, s83, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v33, s61, 1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_writelane_b32 v33, s60, 2
|
|
; VI-NEXT: v_writelane_b32 v33, s59, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB19_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_or_b32 s36, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s5, s9, 16
|
|
; VI-NEXT: s_or_b32 s38, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s11, 16
|
|
; VI-NEXT: s_or_b32 s39, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; VI-NEXT: s_or_b32 s37, s40, s41
|
|
; VI-NEXT: s_or_b32 s40, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; VI-NEXT: s_or_b32 s41, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; VI-NEXT: s_or_b32 s42, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; VI-NEXT: s_or_b32 s43, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s44, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s45, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s46, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s47, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_or_b32 s48, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s5, s56, 16
|
|
; VI-NEXT: s_or_b32 s49, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s83
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s50, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s80
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s51, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s69
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s52, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s34
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s53, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s91
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s54, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s79
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s55, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s76
|
|
; VI-NEXT: s_lshl_b32 s5, s62, 16
|
|
; VI-NEXT: s_mov_b32 s77, s56
|
|
; VI-NEXT: s_or_b32 s56, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s74
|
|
; VI-NEXT: s_lshl_b32 s5, s57, 16
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: s_mov_b32 s75, s57
|
|
; VI-NEXT: s_or_b32 s57, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s72
|
|
; VI-NEXT: s_lshl_b32 s5, s58, 16
|
|
; VI-NEXT: s_mov_b32 s78, s58
|
|
; VI-NEXT: s_or_b32 s58, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s5, s59, 16
|
|
; VI-NEXT: s_or_b32 s59, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s5, s60, 16
|
|
; VI-NEXT: s_or_b32 s60, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s5, s61, 16
|
|
; VI-NEXT: s_or_b32 s61, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_mov_b32 s88, s62
|
|
; VI-NEXT: s_or_b32 s62, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s5, s63, 16
|
|
; VI-NEXT: s_mov_b32 s15, s63
|
|
; VI-NEXT: s_or_b32 s63, s4, s5
|
|
; VI-NEXT: s_cbranch_execnz .LBB19_4
|
|
; VI-NEXT: .LBB19_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; VI-NEXT: v_mov_b32_e32 v0, s13
|
|
; VI-NEXT: v_mov_b32_e32 v2, s11
|
|
; VI-NEXT: v_add_f16_sdwa v0, v0, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, s16, v27
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s17, v27
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v2, s9
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s18, v27
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v3, s7
|
|
; VI-NEXT: v_add_f16_sdwa v3, v3, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, s19, v27
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_mov_b32_e32 v4, s87
|
|
; VI-NEXT: v_add_f16_sdwa v4, v4, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, s20, v27
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_mov_b32_e32 v5, s86
|
|
; VI-NEXT: v_add_f16_sdwa v5, v5, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, s21, v27
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_mov_b32_e32 v6, s84
|
|
; VI-NEXT: v_add_f16_sdwa v6, v6, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, s22, v27
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_mov_b32_e32 v7, s81
|
|
; VI-NEXT: v_add_f16_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, s23, v27
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_mov_b32_e32 v8, s70
|
|
; VI-NEXT: v_add_f16_sdwa v8, v8, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, s24, v27
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_mov_b32_e32 v9, s68
|
|
; VI-NEXT: v_add_f16_sdwa v9, v9, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, s25, v27
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_mov_b32_e32 v10, s31
|
|
; VI-NEXT: v_add_f16_sdwa v10, v10, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, s26, v27
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_mov_b32_e32 v11, s90
|
|
; VI-NEXT: v_add_f16_sdwa v11, v11, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, s27, v27
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_mov_b32_e32 v12, s73
|
|
; VI-NEXT: v_add_f16_sdwa v12, v12, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, s28, v27
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_mov_b32_e32 v13, s77
|
|
; VI-NEXT: v_add_f16_sdwa v13, v13, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, s29, v27
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_mov_b32_e32 v14, s85
|
|
; VI-NEXT: v_add_f16_sdwa v14, v14, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, s83, v27
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_mov_b32_e32 v15, s82
|
|
; VI-NEXT: v_add_f16_sdwa v15, v15, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, s80, v27
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_mov_b32_e32 v16, s71
|
|
; VI-NEXT: v_add_f16_sdwa v16, v16, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, s69, v27
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_mov_b32_e32 v17, s35
|
|
; VI-NEXT: v_add_f16_sdwa v17, v17, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, s34, v27
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_mov_b32_e32 v18, s30
|
|
; VI-NEXT: v_add_f16_sdwa v18, v18, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, s91, v27
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_mov_b32_e32 v19, s89
|
|
; VI-NEXT: v_add_f16_sdwa v19, v19, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, s79, v27
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_mov_b32_e32 v20, s88
|
|
; VI-NEXT: v_add_f16_sdwa v20, v20, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, s76, v27
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_mov_b32_e32 v21, s75
|
|
; VI-NEXT: v_add_f16_sdwa v21, v21, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, s74, v27
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_mov_b32_e32 v22, s78
|
|
; VI-NEXT: v_add_f16_sdwa v22, v22, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v23, s72, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 3
|
|
; VI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; VI-NEXT: v_mov_b32_e32 v23, s4
|
|
; VI-NEXT: v_add_f16_sdwa v23, v23, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, s14, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 2
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: v_mov_b32_e32 v24, s4
|
|
; VI-NEXT: v_add_f16_sdwa v24, v24, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v25, s12, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 1
|
|
; VI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; VI-NEXT: v_mov_b32_e32 v25, s4
|
|
; VI-NEXT: v_add_f16_sdwa v25, v25, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v26, s10, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 0
|
|
; VI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; VI-NEXT: v_mov_b32_e32 v26, s4
|
|
; VI-NEXT: v_add_f16_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, s8, v27
|
|
; VI-NEXT: v_or_b32_e32 v26, v28, v26
|
|
; VI-NEXT: v_mov_b32_e32 v28, s15
|
|
; VI-NEXT: v_add_f16_sdwa v28, v28, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v27, s6, v27
|
|
; VI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; VI-NEXT: s_branch .LBB19_5
|
|
; VI-NEXT: .LBB19_3:
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: s_mov_b32 s88, s62
|
|
; VI-NEXT: s_mov_b32 s77, s56
|
|
; VI-NEXT: s_mov_b32 s75, s57
|
|
; VI-NEXT: s_mov_b32 s78, s58
|
|
; VI-NEXT: s_mov_b32 s15, s63
|
|
; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; VI-NEXT: s_branch .LBB19_2
|
|
; VI-NEXT: .LBB19_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_mov_b32_e32 v28, s64
|
|
; VI-NEXT: v_mov_b32_e32 v29, s65
|
|
; VI-NEXT: v_mov_b32_e32 v30, s66
|
|
; VI-NEXT: v_mov_b32_e32 v31, s67
|
|
; VI-NEXT: .LBB19_5: ; %end
|
|
; VI-NEXT: v_readlane_b32 s87, v32, 31
|
|
; VI-NEXT: v_readlane_b32 s86, v32, 30
|
|
; VI-NEXT: v_readlane_b32 s85, v32, 29
|
|
; VI-NEXT: v_readlane_b32 s84, v32, 28
|
|
; VI-NEXT: v_readlane_b32 s83, v32, 27
|
|
; VI-NEXT: v_readlane_b32 s82, v32, 26
|
|
; VI-NEXT: v_readlane_b32 s81, v32, 25
|
|
; VI-NEXT: v_readlane_b32 s80, v32, 24
|
|
; VI-NEXT: v_readlane_b32 s71, v32, 23
|
|
; VI-NEXT: v_readlane_b32 s70, v32, 22
|
|
; VI-NEXT: v_readlane_b32 s69, v32, 21
|
|
; VI-NEXT: v_readlane_b32 s68, v32, 20
|
|
; VI-NEXT: v_readlane_b32 s67, v32, 19
|
|
; VI-NEXT: v_readlane_b32 s66, v32, 18
|
|
; VI-NEXT: v_readlane_b32 s65, v32, 17
|
|
; VI-NEXT: v_readlane_b32 s64, v32, 16
|
|
; VI-NEXT: v_readlane_b32 s55, v32, 15
|
|
; VI-NEXT: v_readlane_b32 s54, v32, 14
|
|
; VI-NEXT: v_readlane_b32 s53, v32, 13
|
|
; VI-NEXT: v_readlane_b32 s52, v32, 12
|
|
; VI-NEXT: v_readlane_b32 s51, v32, 11
|
|
; VI-NEXT: v_readlane_b32 s50, v32, 10
|
|
; VI-NEXT: v_readlane_b32 s49, v32, 9
|
|
; VI-NEXT: v_readlane_b32 s48, v32, 8
|
|
; VI-NEXT: v_readlane_b32 s39, v32, 7
|
|
; VI-NEXT: v_readlane_b32 s38, v32, 6
|
|
; VI-NEXT: v_readlane_b32 s37, v32, 5
|
|
; VI-NEXT: v_readlane_b32 s36, v32, 4
|
|
; VI-NEXT: v_readlane_b32 s35, v32, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v32, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v32, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v32, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56f16_to_v28i32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: s_lshr_b32 s12, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s13, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s14, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s15, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s88, v5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v3
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v2
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: s_lshr_b32 s4, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s5, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s6, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s7, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s8, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s9, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s10, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s63, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s62, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s61, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s60, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s59, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s58, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s57, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s56, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s88, 16
|
|
; GFX9-NEXT: s_lshr_b32 s17, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s18, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s19, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s20, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s21, s12, 16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_cmp_lg_u32 s42, 0
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s50, s12, s21
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s51, s13, s20
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s52, s14, s19
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s53, s15, s18
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s54, s16, s17
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s55, s88, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s56, s56, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s57, s57, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s58, s58, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s59, s59, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s60, s60, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s61, s61, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s62, s62, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s63, s63, s72
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB19_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB19_4
|
|
; GFX9-NEXT: .LBB19_2: ; %cmp.true
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; GFX9-NEXT: v_pk_add_f16 v0, s36, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, s37, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, s38, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, s39, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, s40, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, s41, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, s42, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, s43, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, s44, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v9, s45, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, s46, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, s47, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, s48, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, s49, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, s50, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, s51, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, s52, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, s53, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, s54, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, s55, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, s56, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, s57, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, s58, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, s59, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v24, s60, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v25, s61, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v26, s62, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v27, s63, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_branch .LBB19_5
|
|
; GFX9-NEXT: .LBB19_3:
|
|
; GFX9-NEXT: s_branch .LBB19_2
|
|
; GFX9-NEXT: .LBB19_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB19_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v56f16_to_v28i32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s44, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s74, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s76, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s77, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s78, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s88, v10
|
|
; GFX11-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s0, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s44, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s56, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s58, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s59, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s60, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s62, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s74, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s76, 16
|
|
; GFX11-NEXT: s_lshr_b32 s90, s77, 16
|
|
; GFX11-NEXT: s_lshr_b32 s91, s78, 16
|
|
; GFX11-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s88, 0
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s78, s91
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s77, s90
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s76, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s74, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s62, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s60, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s59, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s58, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s56, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s44, s57
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB19_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB19_4
|
|
; GFX11-NEXT: .LBB19_2: ; %cmp.true
|
|
; GFX11-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v18, 0x200, s18 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v19, 0x200, s19 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v20, 0x200, s20 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v21, 0x200, s21 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v22, 0x200, s22 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v23, 0x200, s23 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v24, 0x200, s24 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v25, 0x200, s25 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v26, 0x200, s26 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v27, 0x200, s27 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB19_3:
|
|
; GFX11-NEXT: s_branch .LBB19_2
|
|
; GFX11-NEXT: .LBB19_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <56 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <56 x half> %a1 to <28 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x half> %a to <28 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x i32> %phi
|
|
}
|
|
|
|
define <14 x i64> @bitcast_v28f32_to_v14i64(<28 x float> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v28f32_to_v14i64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB20_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; SI-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; SI-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; SI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; SI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; SI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; SI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; SI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; SI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; SI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; SI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; SI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; SI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; SI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; SI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; SI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; SI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; SI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; SI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; SI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; SI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; SI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; SI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; SI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; SI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; SI-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; SI-NEXT: .LBB20_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28f32_to_v14i64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB20_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; VI-NEXT: .LBB20_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28f32_to_v14i64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB20_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; GFX9-NEXT: .LBB20_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v28f32_to_v14i64:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB20_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
|
|
; GFX11-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
|
|
; GFX11-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
|
|
; GFX11-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
|
|
; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
|
|
; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
|
|
; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
|
|
; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
|
|
; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
|
|
; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
|
|
; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
|
|
; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
|
|
; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
|
|
; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
|
|
; GFX11-NEXT: .LBB20_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <28 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <28 x float> %a1 to <14 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x float> %a to <14 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x i64> %phi
|
|
}
|
|
|
|
define inreg <14 x i64> @bitcast_v28f32_to_v14i64_scalar(<28 x float> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v28f32_to_v14i64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; SI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; SI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; SI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; SI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; SI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; SI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; SI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; SI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; SI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; SI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; SI-NEXT: s_mov_b32 s49, s29
|
|
; SI-NEXT: s_mov_b32 s48, s28
|
|
; SI-NEXT: s_mov_b32 s47, s27
|
|
; SI-NEXT: s_mov_b32 s46, s26
|
|
; SI-NEXT: s_mov_b32 s45, s25
|
|
; SI-NEXT: s_mov_b32 s44, s24
|
|
; SI-NEXT: s_mov_b32 s43, s23
|
|
; SI-NEXT: s_mov_b32 s42, s22
|
|
; SI-NEXT: s_mov_b32 s41, s21
|
|
; SI-NEXT: s_mov_b32 s40, s20
|
|
; SI-NEXT: s_mov_b32 s39, s19
|
|
; SI-NEXT: s_mov_b32 s38, s18
|
|
; SI-NEXT: s_mov_b32 s37, s17
|
|
; SI-NEXT: s_mov_b32 s36, s16
|
|
; SI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB21_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB21_4
|
|
; SI-NEXT: .LBB21_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; SI-NEXT: s_branch .LBB21_5
|
|
; SI-NEXT: .LBB21_3:
|
|
; SI-NEXT: s_branch .LBB21_2
|
|
; SI-NEXT: .LBB21_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_mov_b32_e32 v28, s64
|
|
; SI-NEXT: v_mov_b32_e32 v29, s65
|
|
; SI-NEXT: v_mov_b32_e32 v30, s66
|
|
; SI-NEXT: v_mov_b32_e32 v31, s67
|
|
; SI-NEXT: .LBB21_5: ; %end
|
|
; SI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; SI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; SI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; SI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; SI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; SI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; SI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; SI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; SI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; SI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; SI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; SI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28f32_to_v14i64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; VI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; VI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; VI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; VI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; VI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; VI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; VI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; VI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; VI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; VI-NEXT: s_mov_b32 s49, s29
|
|
; VI-NEXT: s_mov_b32 s48, s28
|
|
; VI-NEXT: s_mov_b32 s47, s27
|
|
; VI-NEXT: s_mov_b32 s46, s26
|
|
; VI-NEXT: s_mov_b32 s45, s25
|
|
; VI-NEXT: s_mov_b32 s44, s24
|
|
; VI-NEXT: s_mov_b32 s43, s23
|
|
; VI-NEXT: s_mov_b32 s42, s22
|
|
; VI-NEXT: s_mov_b32 s41, s21
|
|
; VI-NEXT: s_mov_b32 s40, s20
|
|
; VI-NEXT: s_mov_b32 s39, s19
|
|
; VI-NEXT: s_mov_b32 s38, s18
|
|
; VI-NEXT: s_mov_b32 s37, s17
|
|
; VI-NEXT: s_mov_b32 s36, s16
|
|
; VI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB21_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB21_4
|
|
; VI-NEXT: .LBB21_2: ; %cmp.true
|
|
; VI-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; VI-NEXT: s_branch .LBB21_5
|
|
; VI-NEXT: .LBB21_3:
|
|
; VI-NEXT: s_branch .LBB21_2
|
|
; VI-NEXT: .LBB21_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_mov_b32_e32 v28, s64
|
|
; VI-NEXT: v_mov_b32_e32 v29, s65
|
|
; VI-NEXT: v_mov_b32_e32 v30, s66
|
|
; VI-NEXT: v_mov_b32_e32 v31, s67
|
|
; VI-NEXT: .LBB21_5: ; %end
|
|
; VI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; VI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; VI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; VI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; VI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; VI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; VI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; VI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; VI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; VI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; VI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; VI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28f32_to_v14i64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_mov_b32 s49, s29
|
|
; GFX9-NEXT: s_mov_b32 s48, s28
|
|
; GFX9-NEXT: s_mov_b32 s47, s27
|
|
; GFX9-NEXT: s_mov_b32 s46, s26
|
|
; GFX9-NEXT: s_mov_b32 s45, s25
|
|
; GFX9-NEXT: s_mov_b32 s44, s24
|
|
; GFX9-NEXT: s_mov_b32 s43, s23
|
|
; GFX9-NEXT: s_mov_b32 s42, s22
|
|
; GFX9-NEXT: s_mov_b32 s41, s21
|
|
; GFX9-NEXT: s_mov_b32 s40, s20
|
|
; GFX9-NEXT: s_mov_b32 s39, s19
|
|
; GFX9-NEXT: s_mov_b32 s38, s18
|
|
; GFX9-NEXT: s_mov_b32 s37, s17
|
|
; GFX9-NEXT: s_mov_b32 s36, s16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s55, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s54, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s53, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s52, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s51, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s50, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB21_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB21_4
|
|
; GFX9-NEXT: .LBB21_2: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; GFX9-NEXT: s_branch .LBB21_5
|
|
; GFX9-NEXT: .LBB21_3:
|
|
; GFX9-NEXT: s_branch .LBB21_2
|
|
; GFX9-NEXT: .LBB21_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB21_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v28f32_to_v14i64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s4, -1
|
|
; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX11-NEXT: s_mov_b32 s36, s0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s0, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s63, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v8
|
|
; GFX11-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s61, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s57, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v2
|
|
; GFX11-NEXT: s_mov_b32 s47, s23
|
|
; GFX11-NEXT: s_mov_b32 s46, s22
|
|
; GFX11-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX11-NEXT: s_mov_b32 s45, s21
|
|
; GFX11-NEXT: s_mov_b32 s44, s20
|
|
; GFX11-NEXT: s_mov_b32 s43, s19
|
|
; GFX11-NEXT: s_mov_b32 s42, s18
|
|
; GFX11-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX11-NEXT: s_mov_b32 s48, s24
|
|
; GFX11-NEXT: s_mov_b32 s41, s17
|
|
; GFX11-NEXT: s_mov_b32 s40, s16
|
|
; GFX11-NEXT: s_mov_b32 s39, s3
|
|
; GFX11-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX11-NEXT: s_mov_b32 s49, s25
|
|
; GFX11-NEXT: s_mov_b32 s38, s2
|
|
; GFX11-NEXT: s_mov_b32 s37, s1
|
|
; GFX11-NEXT: s_cmp_lg_u32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX11-NEXT: s_mov_b32 s50, s26
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX11-NEXT: s_mov_b32 s51, s27
|
|
; GFX11-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX11-NEXT: s_mov_b32 s52, s28
|
|
; GFX11-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX11-NEXT: s_mov_b32 s53, s29
|
|
; GFX11-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s54, v0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX11-NEXT: v_readfirstlane_b32 s55, v1
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB21_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB21_4
|
|
; GFX11-NEXT: .LBB21_2: ; %cmp.true
|
|
; GFX11-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; GFX11-NEXT: s_branch .LBB21_5
|
|
; GFX11-NEXT: .LBB21_3:
|
|
; GFX11-NEXT: s_branch .LBB21_2
|
|
; GFX11-NEXT: .LBB21_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
|
|
; GFX11-NEXT: .LBB21_5: ; %end
|
|
; GFX11-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX11-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX11-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX11-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX11-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX11-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX11-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX11-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX11-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX11-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX11-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX11-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
|
|
; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s0
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <28 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <28 x float> %a1 to <14 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x float> %a to <14 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x i64> %phi
|
|
}
|
|
|
|
define <28 x float> @bitcast_v14i64_to_v28f32(<14 x i64> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v14i64_to_v28f32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB22_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; SI-NEXT: .LBB22_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14i64_to_v28f32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB22_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
|
|
; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
|
|
; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
|
|
; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
|
|
; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
|
|
; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; VI-NEXT: .LBB22_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14i64_to_v28f32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB22_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX9-NEXT: .LBB22_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v14i64_to_v28f32:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB22_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
|
|
; GFX11-NEXT: .LBB22_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <14 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <14 x i64> %a1 to <28 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x i64> %a to <28 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x float> %phi
|
|
}
|
|
|
|
define inreg <28 x float> @bitcast_v14i64_to_v28f32_scalar(<14 x i64> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v14i64_to_v28f32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB23_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB23_3
|
|
; SI-NEXT: .LBB23_2: ; %cmp.true
|
|
; SI-NEXT: s_add_u32 s7, s7, 3
|
|
; SI-NEXT: s_addc_u32 s6, s6, 0
|
|
; SI-NEXT: s_add_u32 s9, s9, 3
|
|
; SI-NEXT: s_addc_u32 s8, s8, 0
|
|
; SI-NEXT: s_add_u32 s11, s11, 3
|
|
; SI-NEXT: s_addc_u32 s10, s10, 0
|
|
; SI-NEXT: s_add_u32 s13, s13, 3
|
|
; SI-NEXT: s_addc_u32 s12, s12, 0
|
|
; SI-NEXT: s_add_u32 s15, s15, 3
|
|
; SI-NEXT: s_addc_u32 s14, s14, 0
|
|
; SI-NEXT: s_add_u32 s41, s41, 3
|
|
; SI-NEXT: s_addc_u32 s40, s40, 0
|
|
; SI-NEXT: s_add_u32 s43, s43, 3
|
|
; SI-NEXT: s_addc_u32 s42, s42, 0
|
|
; SI-NEXT: s_add_u32 s28, s28, 3
|
|
; SI-NEXT: s_addc_u32 s29, s29, 0
|
|
; SI-NEXT: s_add_u32 s26, s26, 3
|
|
; SI-NEXT: s_addc_u32 s27, s27, 0
|
|
; SI-NEXT: s_add_u32 s24, s24, 3
|
|
; SI-NEXT: s_addc_u32 s25, s25, 0
|
|
; SI-NEXT: s_add_u32 s22, s22, 3
|
|
; SI-NEXT: s_addc_u32 s23, s23, 0
|
|
; SI-NEXT: s_add_u32 s20, s20, 3
|
|
; SI-NEXT: s_addc_u32 s21, s21, 0
|
|
; SI-NEXT: s_add_u32 s18, s18, 3
|
|
; SI-NEXT: s_addc_u32 s19, s19, 0
|
|
; SI-NEXT: s_add_u32 s16, s16, 3
|
|
; SI-NEXT: s_addc_u32 s17, s17, 0
|
|
; SI-NEXT: .LBB23_3: ; %end
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s43
|
|
; SI-NEXT: v_mov_b32_e32 v15, s42
|
|
; SI-NEXT: v_mov_b32_e32 v16, s41
|
|
; SI-NEXT: v_mov_b32_e32 v17, s40
|
|
; SI-NEXT: v_mov_b32_e32 v18, s15
|
|
; SI-NEXT: v_mov_b32_e32 v19, s14
|
|
; SI-NEXT: v_mov_b32_e32 v20, s13
|
|
; SI-NEXT: v_mov_b32_e32 v21, s12
|
|
; SI-NEXT: v_mov_b32_e32 v22, s11
|
|
; SI-NEXT: v_mov_b32_e32 v23, s10
|
|
; SI-NEXT: v_mov_b32_e32 v24, s9
|
|
; SI-NEXT: v_mov_b32_e32 v25, s8
|
|
; SI-NEXT: v_mov_b32_e32 v26, s7
|
|
; SI-NEXT: v_mov_b32_e32 v27, s6
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB23_4:
|
|
; SI-NEXT: s_branch .LBB23_2
|
|
;
|
|
; VI-LABEL: bitcast_v14i64_to_v28f32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB23_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB23_3
|
|
; VI-NEXT: .LBB23_2: ; %cmp.true
|
|
; VI-NEXT: s_add_u32 s7, s7, 3
|
|
; VI-NEXT: s_addc_u32 s6, s6, 0
|
|
; VI-NEXT: s_add_u32 s9, s9, 3
|
|
; VI-NEXT: s_addc_u32 s8, s8, 0
|
|
; VI-NEXT: s_add_u32 s11, s11, 3
|
|
; VI-NEXT: s_addc_u32 s10, s10, 0
|
|
; VI-NEXT: s_add_u32 s13, s13, 3
|
|
; VI-NEXT: s_addc_u32 s12, s12, 0
|
|
; VI-NEXT: s_add_u32 s15, s15, 3
|
|
; VI-NEXT: s_addc_u32 s14, s14, 0
|
|
; VI-NEXT: s_add_u32 s41, s41, 3
|
|
; VI-NEXT: s_addc_u32 s40, s40, 0
|
|
; VI-NEXT: s_add_u32 s43, s43, 3
|
|
; VI-NEXT: s_addc_u32 s42, s42, 0
|
|
; VI-NEXT: s_add_u32 s28, s28, 3
|
|
; VI-NEXT: s_addc_u32 s29, s29, 0
|
|
; VI-NEXT: s_add_u32 s26, s26, 3
|
|
; VI-NEXT: s_addc_u32 s27, s27, 0
|
|
; VI-NEXT: s_add_u32 s24, s24, 3
|
|
; VI-NEXT: s_addc_u32 s25, s25, 0
|
|
; VI-NEXT: s_add_u32 s22, s22, 3
|
|
; VI-NEXT: s_addc_u32 s23, s23, 0
|
|
; VI-NEXT: s_add_u32 s20, s20, 3
|
|
; VI-NEXT: s_addc_u32 s21, s21, 0
|
|
; VI-NEXT: s_add_u32 s18, s18, 3
|
|
; VI-NEXT: s_addc_u32 s19, s19, 0
|
|
; VI-NEXT: s_add_u32 s16, s16, 3
|
|
; VI-NEXT: s_addc_u32 s17, s17, 0
|
|
; VI-NEXT: .LBB23_3: ; %end
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v14, s43
|
|
; VI-NEXT: v_mov_b32_e32 v15, s42
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB23_4:
|
|
; VI-NEXT: s_branch .LBB23_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v14i64_to_v28f32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB23_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB23_3
|
|
; GFX9-NEXT: .LBB23_2: ; %cmp.true
|
|
; GFX9-NEXT: s_add_u32 s7, s7, 3
|
|
; GFX9-NEXT: s_addc_u32 s6, s6, 0
|
|
; GFX9-NEXT: s_add_u32 s9, s9, 3
|
|
; GFX9-NEXT: s_addc_u32 s8, s8, 0
|
|
; GFX9-NEXT: s_add_u32 s11, s11, 3
|
|
; GFX9-NEXT: s_addc_u32 s10, s10, 0
|
|
; GFX9-NEXT: s_add_u32 s13, s13, 3
|
|
; GFX9-NEXT: s_addc_u32 s12, s12, 0
|
|
; GFX9-NEXT: s_add_u32 s15, s15, 3
|
|
; GFX9-NEXT: s_addc_u32 s14, s14, 0
|
|
; GFX9-NEXT: s_add_u32 s41, s41, 3
|
|
; GFX9-NEXT: s_addc_u32 s40, s40, 0
|
|
; GFX9-NEXT: s_add_u32 s43, s43, 3
|
|
; GFX9-NEXT: s_addc_u32 s42, s42, 0
|
|
; GFX9-NEXT: s_add_u32 s28, s28, 3
|
|
; GFX9-NEXT: s_addc_u32 s29, s29, 0
|
|
; GFX9-NEXT: s_add_u32 s26, s26, 3
|
|
; GFX9-NEXT: s_addc_u32 s27, s27, 0
|
|
; GFX9-NEXT: s_add_u32 s24, s24, 3
|
|
; GFX9-NEXT: s_addc_u32 s25, s25, 0
|
|
; GFX9-NEXT: s_add_u32 s22, s22, 3
|
|
; GFX9-NEXT: s_addc_u32 s23, s23, 0
|
|
; GFX9-NEXT: s_add_u32 s20, s20, 3
|
|
; GFX9-NEXT: s_addc_u32 s21, s21, 0
|
|
; GFX9-NEXT: s_add_u32 s18, s18, 3
|
|
; GFX9-NEXT: s_addc_u32 s19, s19, 0
|
|
; GFX9-NEXT: s_add_u32 s16, s16, 3
|
|
; GFX9-NEXT: s_addc_u32 s17, s17, 0
|
|
; GFX9-NEXT: .LBB23_3: ; %end
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB23_4:
|
|
; GFX9-NEXT: s_branch .LBB23_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v14i64_to_v28f32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB23_4
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB23_3
|
|
; GFX11-NEXT: .LBB23_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_u32 s5, s5, 3
|
|
; GFX11-NEXT: s_addc_u32 s4, s4, 0
|
|
; GFX11-NEXT: s_add_u32 s7, s7, 3
|
|
; GFX11-NEXT: s_addc_u32 s6, s6, 0
|
|
; GFX11-NEXT: s_add_u32 s9, s9, 3
|
|
; GFX11-NEXT: s_addc_u32 s8, s8, 0
|
|
; GFX11-NEXT: s_add_u32 s11, s11, 3
|
|
; GFX11-NEXT: s_addc_u32 s10, s10, 0
|
|
; GFX11-NEXT: s_add_u32 s13, s13, 3
|
|
; GFX11-NEXT: s_addc_u32 s12, s12, 0
|
|
; GFX11-NEXT: s_add_u32 s28, s28, 3
|
|
; GFX11-NEXT: s_addc_u32 s29, s29, 0
|
|
; GFX11-NEXT: s_add_u32 s26, s26, 3
|
|
; GFX11-NEXT: s_addc_u32 s27, s27, 0
|
|
; GFX11-NEXT: s_add_u32 s24, s24, 3
|
|
; GFX11-NEXT: s_addc_u32 s25, s25, 0
|
|
; GFX11-NEXT: s_add_u32 s22, s22, 3
|
|
; GFX11-NEXT: s_addc_u32 s23, s23, 0
|
|
; GFX11-NEXT: s_add_u32 s20, s20, 3
|
|
; GFX11-NEXT: s_addc_u32 s21, s21, 0
|
|
; GFX11-NEXT: s_add_u32 s18, s18, 3
|
|
; GFX11-NEXT: s_addc_u32 s19, s19, 0
|
|
; GFX11-NEXT: s_add_u32 s16, s16, 3
|
|
; GFX11-NEXT: s_addc_u32 s17, s17, 0
|
|
; GFX11-NEXT: s_add_u32 s2, s2, 3
|
|
; GFX11-NEXT: s_addc_u32 s3, s3, 0
|
|
; GFX11-NEXT: s_add_u32 s0, s0, 3
|
|
; GFX11-NEXT: s_addc_u32 s1, s1, 0
|
|
; GFX11-NEXT: .LBB23_3: ; %end
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB23_4:
|
|
; GFX11-NEXT: s_branch .LBB23_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <14 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <14 x i64> %a1 to <28 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x i64> %a to <28 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x float> %phi
|
|
}
|
|
|
|
define <14 x double> @bitcast_v28f32_to_v14f64(<28 x float> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v28f32_to_v14f64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB24_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; SI-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; SI-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; SI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; SI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; SI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; SI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; SI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; SI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; SI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; SI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; SI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; SI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; SI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; SI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; SI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; SI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; SI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; SI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; SI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; SI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; SI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; SI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; SI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; SI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; SI-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; SI-NEXT: .LBB24_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28f32_to_v14f64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB24_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; VI-NEXT: .LBB24_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28f32_to_v14f64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB24_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; GFX9-NEXT: .LBB24_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v28f32_to_v14f64:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB24_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
|
|
; GFX11-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
|
|
; GFX11-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
|
|
; GFX11-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
|
|
; GFX11-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
|
|
; GFX11-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
|
|
; GFX11-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
|
|
; GFX11-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
|
|
; GFX11-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
|
|
; GFX11-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
|
|
; GFX11-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
|
|
; GFX11-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
|
|
; GFX11-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
|
|
; GFX11-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
|
|
; GFX11-NEXT: .LBB24_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <28 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <28 x float> %a1 to <14 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x float> %a to <14 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x double> %phi
|
|
}
|
|
|
|
define inreg <14 x double> @bitcast_v28f32_to_v14f64_scalar(<28 x float> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v28f32_to_v14f64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; SI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; SI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; SI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; SI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; SI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; SI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; SI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; SI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; SI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; SI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; SI-NEXT: s_mov_b32 s49, s29
|
|
; SI-NEXT: s_mov_b32 s48, s28
|
|
; SI-NEXT: s_mov_b32 s47, s27
|
|
; SI-NEXT: s_mov_b32 s46, s26
|
|
; SI-NEXT: s_mov_b32 s45, s25
|
|
; SI-NEXT: s_mov_b32 s44, s24
|
|
; SI-NEXT: s_mov_b32 s43, s23
|
|
; SI-NEXT: s_mov_b32 s42, s22
|
|
; SI-NEXT: s_mov_b32 s41, s21
|
|
; SI-NEXT: s_mov_b32 s40, s20
|
|
; SI-NEXT: s_mov_b32 s39, s19
|
|
; SI-NEXT: s_mov_b32 s38, s18
|
|
; SI-NEXT: s_mov_b32 s37, s17
|
|
; SI-NEXT: s_mov_b32 s36, s16
|
|
; SI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB25_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB25_4
|
|
; SI-NEXT: .LBB25_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; SI-NEXT: s_branch .LBB25_5
|
|
; SI-NEXT: .LBB25_3:
|
|
; SI-NEXT: s_branch .LBB25_2
|
|
; SI-NEXT: .LBB25_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_mov_b32_e32 v28, s64
|
|
; SI-NEXT: v_mov_b32_e32 v29, s65
|
|
; SI-NEXT: v_mov_b32_e32 v30, s66
|
|
; SI-NEXT: v_mov_b32_e32 v31, s67
|
|
; SI-NEXT: .LBB25_5: ; %end
|
|
; SI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; SI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; SI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; SI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; SI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; SI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; SI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; SI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; SI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; SI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; SI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; SI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28f32_to_v14f64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; VI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; VI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; VI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; VI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; VI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; VI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; VI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; VI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; VI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; VI-NEXT: s_mov_b32 s49, s29
|
|
; VI-NEXT: s_mov_b32 s48, s28
|
|
; VI-NEXT: s_mov_b32 s47, s27
|
|
; VI-NEXT: s_mov_b32 s46, s26
|
|
; VI-NEXT: s_mov_b32 s45, s25
|
|
; VI-NEXT: s_mov_b32 s44, s24
|
|
; VI-NEXT: s_mov_b32 s43, s23
|
|
; VI-NEXT: s_mov_b32 s42, s22
|
|
; VI-NEXT: s_mov_b32 s41, s21
|
|
; VI-NEXT: s_mov_b32 s40, s20
|
|
; VI-NEXT: s_mov_b32 s39, s19
|
|
; VI-NEXT: s_mov_b32 s38, s18
|
|
; VI-NEXT: s_mov_b32 s37, s17
|
|
; VI-NEXT: s_mov_b32 s36, s16
|
|
; VI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB25_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB25_4
|
|
; VI-NEXT: .LBB25_2: ; %cmp.true
|
|
; VI-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; VI-NEXT: s_branch .LBB25_5
|
|
; VI-NEXT: .LBB25_3:
|
|
; VI-NEXT: s_branch .LBB25_2
|
|
; VI-NEXT: .LBB25_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_mov_b32_e32 v28, s64
|
|
; VI-NEXT: v_mov_b32_e32 v29, s65
|
|
; VI-NEXT: v_mov_b32_e32 v30, s66
|
|
; VI-NEXT: v_mov_b32_e32 v31, s67
|
|
; VI-NEXT: .LBB25_5: ; %end
|
|
; VI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; VI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; VI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; VI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; VI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; VI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; VI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; VI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; VI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; VI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; VI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; VI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28f32_to_v14f64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_mov_b32 s49, s29
|
|
; GFX9-NEXT: s_mov_b32 s48, s28
|
|
; GFX9-NEXT: s_mov_b32 s47, s27
|
|
; GFX9-NEXT: s_mov_b32 s46, s26
|
|
; GFX9-NEXT: s_mov_b32 s45, s25
|
|
; GFX9-NEXT: s_mov_b32 s44, s24
|
|
; GFX9-NEXT: s_mov_b32 s43, s23
|
|
; GFX9-NEXT: s_mov_b32 s42, s22
|
|
; GFX9-NEXT: s_mov_b32 s41, s21
|
|
; GFX9-NEXT: s_mov_b32 s40, s20
|
|
; GFX9-NEXT: s_mov_b32 s39, s19
|
|
; GFX9-NEXT: s_mov_b32 s38, s18
|
|
; GFX9-NEXT: s_mov_b32 s37, s17
|
|
; GFX9-NEXT: s_mov_b32 s36, s16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s55, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s54, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s53, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s52, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s51, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s50, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB25_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB25_4
|
|
; GFX9-NEXT: .LBB25_2: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; GFX9-NEXT: s_branch .LBB25_5
|
|
; GFX9-NEXT: .LBB25_3:
|
|
; GFX9-NEXT: s_branch .LBB25_2
|
|
; GFX9-NEXT: .LBB25_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB25_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v28f32_to_v14f64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s4, -1
|
|
; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX11-NEXT: s_mov_b32 s36, s0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s0, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s63, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v8
|
|
; GFX11-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s61, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s57, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v2
|
|
; GFX11-NEXT: s_mov_b32 s47, s23
|
|
; GFX11-NEXT: s_mov_b32 s46, s22
|
|
; GFX11-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX11-NEXT: s_mov_b32 s45, s21
|
|
; GFX11-NEXT: s_mov_b32 s44, s20
|
|
; GFX11-NEXT: s_mov_b32 s43, s19
|
|
; GFX11-NEXT: s_mov_b32 s42, s18
|
|
; GFX11-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX11-NEXT: s_mov_b32 s48, s24
|
|
; GFX11-NEXT: s_mov_b32 s41, s17
|
|
; GFX11-NEXT: s_mov_b32 s40, s16
|
|
; GFX11-NEXT: s_mov_b32 s39, s3
|
|
; GFX11-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX11-NEXT: s_mov_b32 s49, s25
|
|
; GFX11-NEXT: s_mov_b32 s38, s2
|
|
; GFX11-NEXT: s_mov_b32 s37, s1
|
|
; GFX11-NEXT: s_cmp_lg_u32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX11-NEXT: s_mov_b32 s50, s26
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX11-NEXT: s_mov_b32 s51, s27
|
|
; GFX11-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX11-NEXT: s_mov_b32 s52, s28
|
|
; GFX11-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX11-NEXT: s_mov_b32 s53, s29
|
|
; GFX11-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s54, v0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX11-NEXT: v_readfirstlane_b32 s55, v1
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB25_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB25_4
|
|
; GFX11-NEXT: .LBB25_2: ; %cmp.true
|
|
; GFX11-NEXT: v_add_f32_e64 v27, s63, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v26, s62, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v25, s61, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v24, s60, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v23, s59, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v22, s58, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v21, s57, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v20, s56, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v19, s55, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v18, s54, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v17, s53, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v16, s52, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v15, s51, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v14, s50, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v13, s49, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v12, s48, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v11, s47, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v10, s46, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v9, s45, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v8, s44, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v7, s43, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v6, s42, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v5, s41, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v4, s40, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v3, s39, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v2, s38, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v1, s37, 1.0
|
|
; GFX11-NEXT: v_add_f32_e64 v0, s36, 1.0
|
|
; GFX11-NEXT: s_branch .LBB25_5
|
|
; GFX11-NEXT: .LBB25_3:
|
|
; GFX11-NEXT: s_branch .LBB25_2
|
|
; GFX11-NEXT: .LBB25_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
|
|
; GFX11-NEXT: .LBB25_5: ; %end
|
|
; GFX11-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX11-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX11-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX11-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX11-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX11-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX11-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX11-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX11-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX11-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX11-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX11-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
|
|
; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s0
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <28 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <28 x float> %a1 to <14 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x float> %a to <14 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x double> %phi
|
|
}
|
|
|
|
define <28 x float> @bitcast_v14f64_to_v28f32(<14 x double> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v14f64_to_v28f32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB26_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; SI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; SI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; SI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; SI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; SI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; SI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; SI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; SI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; SI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; SI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; SI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; SI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; SI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; SI-NEXT: .LBB26_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14f64_to_v28f32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB26_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; VI-NEXT: .LBB26_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14f64_to_v28f32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB26_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX9-NEXT: .LBB26_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v14f64_to_v28f32:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB26_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX11-NEXT: .LBB26_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <14 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <14 x double> %a1 to <28 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x double> %a to <28 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x float> %phi
|
|
}
|
|
|
|
define inreg <28 x float> @bitcast_v14f64_to_v28f32_scalar(<14 x double> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v14f64_to_v28f32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; SI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; SI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; SI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; SI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; SI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; SI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; SI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; SI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; SI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; SI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; SI-NEXT: s_mov_b32 s49, s29
|
|
; SI-NEXT: s_mov_b32 s48, s28
|
|
; SI-NEXT: s_mov_b32 s47, s27
|
|
; SI-NEXT: s_mov_b32 s46, s26
|
|
; SI-NEXT: s_mov_b32 s45, s25
|
|
; SI-NEXT: s_mov_b32 s44, s24
|
|
; SI-NEXT: s_mov_b32 s43, s23
|
|
; SI-NEXT: s_mov_b32 s42, s22
|
|
; SI-NEXT: s_mov_b32 s41, s21
|
|
; SI-NEXT: s_mov_b32 s40, s20
|
|
; SI-NEXT: s_mov_b32 s39, s19
|
|
; SI-NEXT: s_mov_b32 s38, s18
|
|
; SI-NEXT: s_mov_b32 s37, s17
|
|
; SI-NEXT: s_mov_b32 s36, s16
|
|
; SI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB27_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB27_4
|
|
; SI-NEXT: .LBB27_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; SI-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; SI-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; SI-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; SI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; SI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; SI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; SI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; SI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; SI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; SI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; SI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; SI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; SI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; SI-NEXT: s_branch .LBB27_5
|
|
; SI-NEXT: .LBB27_3:
|
|
; SI-NEXT: s_branch .LBB27_2
|
|
; SI-NEXT: .LBB27_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_mov_b32_e32 v28, s64
|
|
; SI-NEXT: v_mov_b32_e32 v29, s65
|
|
; SI-NEXT: v_mov_b32_e32 v30, s66
|
|
; SI-NEXT: v_mov_b32_e32 v31, s67
|
|
; SI-NEXT: .LBB27_5: ; %end
|
|
; SI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; SI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; SI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; SI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; SI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; SI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; SI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; SI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; SI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; SI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; SI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; SI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14f64_to_v28f32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; VI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; VI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; VI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; VI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; VI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; VI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; VI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; VI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; VI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; VI-NEXT: s_mov_b32 s49, s29
|
|
; VI-NEXT: s_mov_b32 s48, s28
|
|
; VI-NEXT: s_mov_b32 s47, s27
|
|
; VI-NEXT: s_mov_b32 s46, s26
|
|
; VI-NEXT: s_mov_b32 s45, s25
|
|
; VI-NEXT: s_mov_b32 s44, s24
|
|
; VI-NEXT: s_mov_b32 s43, s23
|
|
; VI-NEXT: s_mov_b32 s42, s22
|
|
; VI-NEXT: s_mov_b32 s41, s21
|
|
; VI-NEXT: s_mov_b32 s40, s20
|
|
; VI-NEXT: s_mov_b32 s39, s19
|
|
; VI-NEXT: s_mov_b32 s38, s18
|
|
; VI-NEXT: s_mov_b32 s37, s17
|
|
; VI-NEXT: s_mov_b32 s36, s16
|
|
; VI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB27_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB27_4
|
|
; VI-NEXT: .LBB27_2: ; %cmp.true
|
|
; VI-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; VI-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; VI-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; VI-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; VI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; VI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; VI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; VI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; VI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; VI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; VI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; VI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; VI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; VI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; VI-NEXT: s_branch .LBB27_5
|
|
; VI-NEXT: .LBB27_3:
|
|
; VI-NEXT: s_branch .LBB27_2
|
|
; VI-NEXT: .LBB27_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_mov_b32_e32 v28, s64
|
|
; VI-NEXT: v_mov_b32_e32 v29, s65
|
|
; VI-NEXT: v_mov_b32_e32 v30, s66
|
|
; VI-NEXT: v_mov_b32_e32 v31, s67
|
|
; VI-NEXT: .LBB27_5: ; %end
|
|
; VI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; VI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; VI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; VI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; VI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; VI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; VI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; VI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; VI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; VI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; VI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; VI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14f64_to_v28f32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_mov_b32 s49, s29
|
|
; GFX9-NEXT: s_mov_b32 s48, s28
|
|
; GFX9-NEXT: s_mov_b32 s47, s27
|
|
; GFX9-NEXT: s_mov_b32 s46, s26
|
|
; GFX9-NEXT: s_mov_b32 s45, s25
|
|
; GFX9-NEXT: s_mov_b32 s44, s24
|
|
; GFX9-NEXT: s_mov_b32 s43, s23
|
|
; GFX9-NEXT: s_mov_b32 s42, s22
|
|
; GFX9-NEXT: s_mov_b32 s41, s21
|
|
; GFX9-NEXT: s_mov_b32 s40, s20
|
|
; GFX9-NEXT: s_mov_b32 s39, s19
|
|
; GFX9-NEXT: s_mov_b32 s38, s18
|
|
; GFX9-NEXT: s_mov_b32 s37, s17
|
|
; GFX9-NEXT: s_mov_b32 s36, s16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s55, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s54, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s53, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s52, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s51, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s50, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB27_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB27_4
|
|
; GFX9-NEXT: .LBB27_2: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; GFX9-NEXT: s_branch .LBB27_5
|
|
; GFX9-NEXT: .LBB27_3:
|
|
; GFX9-NEXT: s_branch .LBB27_2
|
|
; GFX9-NEXT: .LBB27_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB27_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v14f64_to_v28f32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s4, -1
|
|
; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX11-NEXT: s_mov_b32 s36, s0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s0, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s63, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v8
|
|
; GFX11-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s61, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s57, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v2
|
|
; GFX11-NEXT: s_mov_b32 s47, s23
|
|
; GFX11-NEXT: s_mov_b32 s46, s22
|
|
; GFX11-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX11-NEXT: s_mov_b32 s45, s21
|
|
; GFX11-NEXT: s_mov_b32 s44, s20
|
|
; GFX11-NEXT: s_mov_b32 s43, s19
|
|
; GFX11-NEXT: s_mov_b32 s42, s18
|
|
; GFX11-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX11-NEXT: s_mov_b32 s48, s24
|
|
; GFX11-NEXT: s_mov_b32 s41, s17
|
|
; GFX11-NEXT: s_mov_b32 s40, s16
|
|
; GFX11-NEXT: s_mov_b32 s39, s3
|
|
; GFX11-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX11-NEXT: s_mov_b32 s49, s25
|
|
; GFX11-NEXT: s_mov_b32 s38, s2
|
|
; GFX11-NEXT: s_mov_b32 s37, s1
|
|
; GFX11-NEXT: s_cmp_lg_u32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX11-NEXT: s_mov_b32 s50, s26
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX11-NEXT: s_mov_b32 s51, s27
|
|
; GFX11-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX11-NEXT: s_mov_b32 s52, s28
|
|
; GFX11-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX11-NEXT: s_mov_b32 s53, s29
|
|
; GFX11-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s54, v0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX11-NEXT: v_readfirstlane_b32 s55, v1
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB27_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB27_4
|
|
; GFX11-NEXT: .LBB27_2: ; %cmp.true
|
|
; GFX11-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; GFX11-NEXT: s_branch .LBB27_5
|
|
; GFX11-NEXT: .LBB27_3:
|
|
; GFX11-NEXT: s_branch .LBB27_2
|
|
; GFX11-NEXT: .LBB27_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
|
|
; GFX11-NEXT: .LBB27_5: ; %end
|
|
; GFX11-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX11-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX11-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX11-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX11-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX11-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX11-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX11-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX11-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX11-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX11-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX11-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
|
|
; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s0
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <14 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <14 x double> %a1 to <28 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x double> %a to <28 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x float> %phi
|
|
}
|
|
|
|
define <56 x i16> @bitcast_v28f32_to_v56i16(<28 x float> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v28f32_to_v56i16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB28_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB28_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB28_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; SI-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; SI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; SI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; SI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; SI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; SI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; SI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; SI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; SI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; SI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; SI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; SI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; SI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; SI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; SI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; SI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; SI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; SI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; SI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; SI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; SI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; SI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; SI-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; SI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; SI-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; SI-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB28_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v47
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v40
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v48
|
|
; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v41
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v53
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v51
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v48
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v38
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28f32_to_v56i16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB28_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB28_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB28_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB28_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28f32_to_v56i16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB28_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB28_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB28_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB28_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v47, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v46, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v45, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v44, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v43, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v42, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v41, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v40, v7, s4
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v8, v55, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v54, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v53, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v52, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v51, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v49, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v48, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v39, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v38, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v37, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v36, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v35, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v34, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v33, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v32, v23, s4
|
|
; GFX9-NEXT: v_perm_b32 v24, v31, v24, s4
|
|
; GFX9-NEXT: v_perm_b32 v25, v30, v25, s4
|
|
; GFX9-NEXT: v_perm_b32 v26, v29, v26, s4
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s4
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v28f32_to_v56i16:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB28_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
|
|
; GFX11-TRUE16-NEXT: .LBB28_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v28f32_to_v56i16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr67
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB28_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB28_2: ; %Flow
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB28_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB28_4: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v71, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v69, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v68, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v67, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v66, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v65, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v64, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v55, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v54, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v53, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v52, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v51, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v49, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v39, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v37, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v36, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v35, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v34, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v33, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v32, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v31, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v30, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v29, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <28 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <28 x float> %a1 to <56 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x float> %a to <56 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x i16> %phi
|
|
}
|
|
|
|
define inreg <56 x i16> @bitcast_v28f32_to_v56i16_scalar(<28 x float> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v28f32_to_v56i16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v58, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v58, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v58, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v58, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v58, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v58, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v58, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v58, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v58, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v58, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v58, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v58, s51, 11
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v14
|
|
; SI-NEXT: v_writelane_b32 v58, s52, 12
|
|
; SI-NEXT: v_readfirstlane_b32 s5, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s40, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v0
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_writelane_b32 v58, s53, 13
|
|
; SI-NEXT: s_cbranch_scc0 .LBB29_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s53, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB29_4
|
|
; SI-NEXT: .LBB29_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e64 v27, s5, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v26, s4, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v25, s7, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v24, s6, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[28:29], v[26:27], 16
|
|
; SI-NEXT: v_add_f32_e64 v23, s9, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v22, s8, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[29:30], v[24:25], 16
|
|
; SI-NEXT: v_add_f32_e64 v21, s11, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v20, s10, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[30:31], v[22:23], 16
|
|
; SI-NEXT: v_add_f32_e64 v19, s13, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v18, s12, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[31:32], v[20:21], 16
|
|
; SI-NEXT: v_add_f32_e64 v17, s15, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v16, s14, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[32:33], v[18:19], 16
|
|
; SI-NEXT: v_add_f32_e64 v15, s41, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v14, s40, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[33:34], v[16:17], 16
|
|
; SI-NEXT: v_add_f32_e64 v13, s29, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v12, s28, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[34:35], v[14:15], 16
|
|
; SI-NEXT: v_add_f32_e64 v5, s21, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v4, s20, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v11, s27, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v10, s26, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[35:36], v[12:13], 16
|
|
; SI-NEXT: v_add_f32_e64 v3, s19, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v2, s18, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v9, s25, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v8, s24, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[36:37], v[10:11], 16
|
|
; SI-NEXT: v_lshr_b64 v[48:49], v[4:5], 16
|
|
; SI-NEXT: v_add_f32_e64 v1, s17, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v0, s16, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v7, s23, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v6, s22, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[37:38], v[8:9], 16
|
|
; SI-NEXT: v_lshr_b64 v[49:50], v[2:3], 16
|
|
; SI-NEXT: v_lshr_b64 v[38:39], v[6:7], 16
|
|
; SI-NEXT: v_lshr_b64 v[50:51], v[0:1], 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v1
|
|
; SI-NEXT: s_branch .LBB29_5
|
|
; SI-NEXT: .LBB29_3:
|
|
; SI-NEXT: ; implicit-def: $sgpr92
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: ; implicit-def: $sgpr31
|
|
; SI-NEXT: ; implicit-def: $sgpr88
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr78
|
|
; SI-NEXT: ; implicit-def: $sgpr35
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr36
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr37
|
|
; SI-NEXT: ; implicit-def: $sgpr72
|
|
; SI-NEXT: ; implicit-def: $sgpr38
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr39
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr48
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr49
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr50
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr51
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr52
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr53
|
|
; SI-NEXT: s_branch .LBB29_2
|
|
; SI-NEXT: .LBB29_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s40
|
|
; SI-NEXT: v_mov_b32_e32 v15, s41
|
|
; SI-NEXT: v_mov_b32_e32 v16, s14
|
|
; SI-NEXT: v_mov_b32_e32 v17, s15
|
|
; SI-NEXT: v_mov_b32_e32 v18, s12
|
|
; SI-NEXT: v_mov_b32_e32 v19, s13
|
|
; SI-NEXT: v_mov_b32_e32 v20, s10
|
|
; SI-NEXT: v_mov_b32_e32 v21, s11
|
|
; SI-NEXT: v_mov_b32_e32 v22, s8
|
|
; SI-NEXT: v_mov_b32_e32 v23, s9
|
|
; SI-NEXT: v_mov_b32_e32 v24, s6
|
|
; SI-NEXT: v_mov_b32_e32 v25, s7
|
|
; SI-NEXT: v_mov_b32_e32 v26, s4
|
|
; SI-NEXT: v_mov_b32_e32 v27, s5
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v57, s30
|
|
; SI-NEXT: v_mov_b32_e32 v56, s31
|
|
; SI-NEXT: v_mov_b32_e32 v47, s34
|
|
; SI-NEXT: v_mov_b32_e32 v46, s35
|
|
; SI-NEXT: v_mov_b32_e32 v45, s36
|
|
; SI-NEXT: v_mov_b32_e32 v44, s37
|
|
; SI-NEXT: v_mov_b32_e32 v43, s38
|
|
; SI-NEXT: v_mov_b32_e32 v42, s39
|
|
; SI-NEXT: v_mov_b32_e32 v41, s48
|
|
; SI-NEXT: v_mov_b32_e32 v40, s49
|
|
; SI-NEXT: v_mov_b32_e32 v55, s50
|
|
; SI-NEXT: v_mov_b32_e32 v54, s51
|
|
; SI-NEXT: v_mov_b32_e32 v53, s52
|
|
; SI-NEXT: v_mov_b32_e32 v52, s53
|
|
; SI-NEXT: v_mov_b32_e32 v28, s42
|
|
; SI-NEXT: v_mov_b32_e32 v29, s44
|
|
; SI-NEXT: v_mov_b32_e32 v30, s46
|
|
; SI-NEXT: v_mov_b32_e32 v31, s56
|
|
; SI-NEXT: v_mov_b32_e32 v32, s58
|
|
; SI-NEXT: v_mov_b32_e32 v33, s60
|
|
; SI-NEXT: v_mov_b32_e32 v34, s62
|
|
; SI-NEXT: v_mov_b32_e32 v35, s72
|
|
; SI-NEXT: v_mov_b32_e32 v36, s74
|
|
; SI-NEXT: v_mov_b32_e32 v37, s76
|
|
; SI-NEXT: v_mov_b32_e32 v38, s78
|
|
; SI-NEXT: v_mov_b32_e32 v48, s88
|
|
; SI-NEXT: v_mov_b32_e32 v49, s90
|
|
; SI-NEXT: v_mov_b32_e32 v50, s92
|
|
; SI-NEXT: .LBB29_5: ; %end
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v50
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v39
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v57
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v49
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v39
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v56
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v48
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v41
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v40
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v39
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v38
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v37
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: v_readlane_b32 s53, v58, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v58, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v58, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v58, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v58, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v58, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v58, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v58, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v58, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v58, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v58, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v58, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v58, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v58, 0
|
|
; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28f32_to_v56i16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v56, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v56, s31, 1
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v56, s34, 2
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_writelane_b32 v56, s35, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB29_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s44, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s42, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB29_4
|
|
; VI-NEXT: .LBB29_2: ; %cmp.true
|
|
; VI-NEXT: v_add_f32_e64 v27, s6, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v26, s7, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v25, s8, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v24, s9, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v23, s10, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v22, s11, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v21, s12, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v20, s13, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v19, s14, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v18, s15, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v17, s40, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v16, s41, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v15, s42, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v14, s43, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v13, s29, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v12, s28, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v11, s27, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v10, s26, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v9, s25, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v8, s24, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v7, s23, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v6, s22, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v5, s21, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v4, s20, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v3, s19, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v2, s18, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v1, s17, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v0, s16, 1.0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: s_branch .LBB29_5
|
|
; VI-NEXT: .LBB29_3:
|
|
; VI-NEXT: ; implicit-def: $sgpr35
|
|
; VI-NEXT: ; implicit-def: $sgpr34
|
|
; VI-NEXT: ; implicit-def: $sgpr31
|
|
; VI-NEXT: ; implicit-def: $sgpr30
|
|
; VI-NEXT: ; implicit-def: $sgpr91
|
|
; VI-NEXT: ; implicit-def: $sgpr90
|
|
; VI-NEXT: ; implicit-def: $sgpr89
|
|
; VI-NEXT: ; implicit-def: $sgpr88
|
|
; VI-NEXT: ; implicit-def: $sgpr79
|
|
; VI-NEXT: ; implicit-def: $sgpr78
|
|
; VI-NEXT: ; implicit-def: $sgpr77
|
|
; VI-NEXT: ; implicit-def: $sgpr76
|
|
; VI-NEXT: ; implicit-def: $sgpr75
|
|
; VI-NEXT: ; implicit-def: $sgpr74
|
|
; VI-NEXT: ; implicit-def: $sgpr73
|
|
; VI-NEXT: ; implicit-def: $sgpr72
|
|
; VI-NEXT: ; implicit-def: $sgpr63
|
|
; VI-NEXT: ; implicit-def: $sgpr62
|
|
; VI-NEXT: ; implicit-def: $sgpr61
|
|
; VI-NEXT: ; implicit-def: $sgpr60
|
|
; VI-NEXT: ; implicit-def: $sgpr59
|
|
; VI-NEXT: ; implicit-def: $sgpr58
|
|
; VI-NEXT: ; implicit-def: $sgpr57
|
|
; VI-NEXT: ; implicit-def: $sgpr56
|
|
; VI-NEXT: ; implicit-def: $sgpr47
|
|
; VI-NEXT: ; implicit-def: $sgpr46
|
|
; VI-NEXT: ; implicit-def: $sgpr45
|
|
; VI-NEXT: ; implicit-def: $sgpr44
|
|
; VI-NEXT: s_branch .LBB29_2
|
|
; VI-NEXT: .LBB29_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v14, s43
|
|
; VI-NEXT: v_mov_b32_e32 v15, s42
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: v_mov_b32_e32 v47, s35
|
|
; VI-NEXT: v_mov_b32_e32 v46, s34
|
|
; VI-NEXT: v_mov_b32_e32 v45, s31
|
|
; VI-NEXT: v_mov_b32_e32 v44, s30
|
|
; VI-NEXT: v_mov_b32_e32 v43, s91
|
|
; VI-NEXT: v_mov_b32_e32 v42, s90
|
|
; VI-NEXT: v_mov_b32_e32 v41, s89
|
|
; VI-NEXT: v_mov_b32_e32 v40, s88
|
|
; VI-NEXT: v_mov_b32_e32 v55, s79
|
|
; VI-NEXT: v_mov_b32_e32 v54, s78
|
|
; VI-NEXT: v_mov_b32_e32 v53, s77
|
|
; VI-NEXT: v_mov_b32_e32 v52, s76
|
|
; VI-NEXT: v_mov_b32_e32 v51, s75
|
|
; VI-NEXT: v_mov_b32_e32 v50, s74
|
|
; VI-NEXT: v_mov_b32_e32 v49, s73
|
|
; VI-NEXT: v_mov_b32_e32 v48, s72
|
|
; VI-NEXT: v_mov_b32_e32 v39, s63
|
|
; VI-NEXT: v_mov_b32_e32 v38, s62
|
|
; VI-NEXT: v_mov_b32_e32 v37, s61
|
|
; VI-NEXT: v_mov_b32_e32 v36, s60
|
|
; VI-NEXT: v_mov_b32_e32 v35, s59
|
|
; VI-NEXT: v_mov_b32_e32 v34, s58
|
|
; VI-NEXT: v_mov_b32_e32 v33, s57
|
|
; VI-NEXT: v_mov_b32_e32 v32, s56
|
|
; VI-NEXT: v_mov_b32_e32 v31, s47
|
|
; VI-NEXT: v_mov_b32_e32 v30, s46
|
|
; VI-NEXT: v_mov_b32_e32 v29, s45
|
|
; VI-NEXT: v_mov_b32_e32 v28, s44
|
|
; VI-NEXT: .LBB29_5: ; %end
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_readlane_b32 s35, v56, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v56, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v56, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v56, 0
|
|
; VI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28f32_to_v56i16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB29_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s44, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s42, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s43, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB29_4
|
|
; GFX9-NEXT: .LBB29_2: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f32_e64 v27, s6, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v26, s7, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v25, s8, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v24, s9, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v23, s10, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v22, s11, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v21, s12, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v20, s13, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v19, s14, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v18, s15, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v17, s40, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v16, s41, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v15, s42, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v14, s43, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v13, s29, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v12, s28, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v11, s27, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v10, s26, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v9, s25, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v8, s24, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v7, s23, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v6, s22, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v5, s21, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v4, s20, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v3, s19, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v2, s18, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v1, s17, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v0, s16, 1.0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: s_branch .LBB29_5
|
|
; GFX9-NEXT: .LBB29_3:
|
|
; GFX9-NEXT: ; implicit-def: $sgpr95
|
|
; GFX9-NEXT: ; implicit-def: $sgpr94
|
|
; GFX9-NEXT: ; implicit-def: $sgpr93
|
|
; GFX9-NEXT: ; implicit-def: $sgpr92
|
|
; GFX9-NEXT: ; implicit-def: $sgpr91
|
|
; GFX9-NEXT: ; implicit-def: $sgpr90
|
|
; GFX9-NEXT: ; implicit-def: $sgpr89
|
|
; GFX9-NEXT: ; implicit-def: $sgpr88
|
|
; GFX9-NEXT: ; implicit-def: $sgpr79
|
|
; GFX9-NEXT: ; implicit-def: $sgpr78
|
|
; GFX9-NEXT: ; implicit-def: $sgpr77
|
|
; GFX9-NEXT: ; implicit-def: $sgpr76
|
|
; GFX9-NEXT: ; implicit-def: $sgpr75
|
|
; GFX9-NEXT: ; implicit-def: $sgpr74
|
|
; GFX9-NEXT: ; implicit-def: $sgpr73
|
|
; GFX9-NEXT: ; implicit-def: $sgpr72
|
|
; GFX9-NEXT: ; implicit-def: $sgpr63
|
|
; GFX9-NEXT: ; implicit-def: $sgpr62
|
|
; GFX9-NEXT: ; implicit-def: $sgpr61
|
|
; GFX9-NEXT: ; implicit-def: $sgpr60
|
|
; GFX9-NEXT: ; implicit-def: $sgpr59
|
|
; GFX9-NEXT: ; implicit-def: $sgpr58
|
|
; GFX9-NEXT: ; implicit-def: $sgpr57
|
|
; GFX9-NEXT: ; implicit-def: $sgpr56
|
|
; GFX9-NEXT: ; implicit-def: $sgpr47
|
|
; GFX9-NEXT: ; implicit-def: $sgpr46
|
|
; GFX9-NEXT: ; implicit-def: $sgpr45
|
|
; GFX9-NEXT: ; implicit-def: $sgpr44
|
|
; GFX9-NEXT: s_branch .LBB29_2
|
|
; GFX9-NEXT: .LBB29_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, s95
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, s94
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, s93
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, s92
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, s91
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, s90
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, s89
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, s88
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, s79
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, s78
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, s77
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, s76
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, s75
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, s74
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, s73
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, s72
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s44
|
|
; GFX9-NEXT: .LBB29_5: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX9-NEXT: v_lshl_or_b32 v0, v47, 16, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v1, v46, 16, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v2, v45, 16, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v3, v44, 16, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v4, v43, 16, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v5, v42, 16, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v6, v41, 16, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v7, v40, 16, v7
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX9-NEXT: v_lshl_or_b32 v8, v55, 16, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v9, v54, 16, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v53, 16, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v52, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v51, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v50, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v49, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v48, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v39, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v38, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v37, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v36, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v35, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v34, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v33, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v32, 16, v23
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v31, 16, v24
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v30, 16, v25
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v29, 16, v26
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v28, 16, v27
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v28f32_to_v56i16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB29_3
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s4, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s5, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s6, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s7, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s8, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s9, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s10, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s12, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s56, s13, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s57, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s58, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s59, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s60, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s61, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s62, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s63, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s72, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s73, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s74, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s75, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s76, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s77, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s78, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s79, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s88, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s89, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s90, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB29_4
|
|
; GFX11-TRUE16-NEXT: .LBB29_2: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v27, s4, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v26, s5, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v25, s6, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v24, s7, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v23, s8, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v22, s9, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v21, s10, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v20, s11, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v19, s12, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v18, s13, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v17, s29, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v16, s28, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v15, s27, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v14, s26, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v13, s25, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v12, s24, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v11, s23, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v10, s22, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v9, s21, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v8, s20, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v7, s19, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v6, s18, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v5, s17, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v4, s16, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v3, s3, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v2, s2, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v1, s1, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v0, s0, 1.0
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB29_5
|
|
; GFX11-TRUE16-NEXT: .LBB29_3:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr90
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB29_2
|
|
; GFX11-TRUE16-NEXT: .LBB29_4:
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v71, s90 :: v_dual_mov_b32 v70, s89
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v69, s88 :: v_dual_mov_b32 v68, s79
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v67, s78 :: v_dual_mov_b32 v66, s77
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v65, s76 :: v_dual_mov_b32 v64, s75
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v55, s74 :: v_dual_mov_b32 v54, s73
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, s72 :: v_dual_mov_b32 v52, s63
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v51, s62 :: v_dual_mov_b32 v50, s61
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v49, s60 :: v_dual_mov_b32 v48, s59
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, s58 :: v_dual_mov_b32 v38, s57
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, s56 :: v_dual_mov_b32 v36, s47
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15
|
|
; GFX11-TRUE16-NEXT: .LBB29_5: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v71.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v70.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v69.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v68.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v67.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v66.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v65.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v64.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v31.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v28.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v28f32_to_v56i16_scalar:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB29_3
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s4, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s5, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s6, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s7, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s8, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s9, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s10, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s12, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s56, s13, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s57, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s58, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s59, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s60, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s61, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s62, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s63, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s72, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s73, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s74, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s75, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s76, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s77, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s78, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s79, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s88, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s89, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s90, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB29_4
|
|
; GFX11-FAKE16-NEXT: .LBB29_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v23, s4, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v24, s5, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v25, s6, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v26, s7, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v27, s8, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v18, s9, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v19, s10, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v20, s11, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v21, s12, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v22, s13, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v13, s29, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v14, s28, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v15, s27, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v16, s26, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v17, s25, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v8, s24, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v9, s23, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v10, s22, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v11, s21, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v12, s20, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v3, s19, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v4, s18, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v5, s17, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v6, s16, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v7, s3, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v0, s2, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v1, s1, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v2, s0, 1.0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB29_5
|
|
; GFX11-FAKE16-NEXT: .LBB29_3:
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr90
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB29_2
|
|
; GFX11-FAKE16-NEXT: .LBB29_4:
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v7, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s18 :: v_dual_mov_b32 v3, s19
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s20 :: v_dual_mov_b32 v11, s21
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v9, s23
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s24 :: v_dual_mov_b32 v17, s25
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s28 :: v_dual_mov_b32 v13, s29
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v21, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v19, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v18, s9 :: v_dual_mov_b32 v27, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v26, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s5 :: v_dual_mov_b32 v23, s4
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v71, s90 :: v_dual_mov_b32 v70, s89
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v69, s88 :: v_dual_mov_b32 v68, s79
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v67, s78 :: v_dual_mov_b32 v66, s77
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v65, s76 :: v_dual_mov_b32 v64, s75
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v55, s74 :: v_dual_mov_b32 v54, s73
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v53, s72 :: v_dual_mov_b32 v52, s63
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v51, s62 :: v_dual_mov_b32 v50, s61
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v49, s60 :: v_dual_mov_b32 v48, s59
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v39, s58 :: v_dual_mov_b32 v38, s57
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, s56 :: v_dual_mov_b32 v36, s47
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15
|
|
; GFX11-FAKE16-NEXT: .LBB29_5: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v80, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v69, 16, v80
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v69, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v70, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v68, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v67, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v65, 16, v69
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v64, 16, v70
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v65, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v55, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v54, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v52, 16, v64
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v51, 16, v65
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v13
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v50, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v49, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v39, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v38, 16, v52
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v19
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v23
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v31, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v29, 16, v33
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v28, 16, v34
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <28 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <28 x float> %a1 to <56 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x float> %a to <56 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x i16> %phi
|
|
}
|
|
|
|
define <28 x float> @bitcast_v56i16_to_v28f32(<56 x i16> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v56i16_to_v28f32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_mov_b32_e32 v48, v19
|
|
; SI-NEXT: v_mov_b32_e32 v49, v18
|
|
; SI-NEXT: v_mov_b32_e32 v50, v17
|
|
; SI-NEXT: v_mov_b32_e32 v51, v16
|
|
; SI-NEXT: v_mov_b32_e32 v52, v15
|
|
; SI-NEXT: v_mov_b32_e32 v53, v14
|
|
; SI-NEXT: v_mov_b32_e32 v54, v13
|
|
; SI-NEXT: v_mov_b32_e32 v55, v12
|
|
; SI-NEXT: v_mov_b32_e32 v40, v11
|
|
; SI-NEXT: v_mov_b32_e32 v41, v10
|
|
; SI-NEXT: v_mov_b32_e32 v42, v9
|
|
; SI-NEXT: v_mov_b32_e32 v43, v8
|
|
; SI-NEXT: v_mov_b32_e32 v44, v7
|
|
; SI-NEXT: v_mov_b32_e32 v45, v6
|
|
; SI-NEXT: v_mov_b32_e32 v46, v5
|
|
; SI-NEXT: v_mov_b32_e32 v47, v4
|
|
; SI-NEXT: v_mov_b32_e32 v56, v3
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_mov_b32_e32 v57, v2
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_mov_b32_e32 v58, v1
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_mov_b32_e32 v59, v0
|
|
; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v26
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v25
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v23
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v22
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v21
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v8, 16, v48
|
|
; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v49
|
|
; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v50
|
|
; SI-NEXT: v_lshrrev_b32_e32 v11, 16, v51
|
|
; SI-NEXT: v_lshrrev_b32_e32 v12, 16, v52
|
|
; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v53
|
|
; SI-NEXT: v_lshrrev_b32_e32 v14, 16, v54
|
|
; SI-NEXT: v_lshrrev_b32_e32 v15, 16, v55
|
|
; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v40
|
|
; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v41
|
|
; SI-NEXT: v_lshrrev_b32_e32 v18, 16, v42
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v43
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v20, 16, v44
|
|
; SI-NEXT: v_lshrrev_b32_e32 v21, 16, v45
|
|
; SI-NEXT: v_lshrrev_b32_e32 v22, 16, v46
|
|
; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v47
|
|
; SI-NEXT: v_lshrrev_b32_e32 v24, 16, v56
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v57
|
|
; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v58
|
|
; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v59
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB30_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v55
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v59
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v58
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v57
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v56
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v47
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v46
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v45
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v44
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v43
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v42
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v41
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v39
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v38
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v62
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v37
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v32
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v36
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v61
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v35
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v60
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v34
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v63
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v33
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v54
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v53
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v52
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v51
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v50
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v49
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v27
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: .LBB30_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB30_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v55
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v59
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v58
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v57
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v56
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v47
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v46
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v45
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v44
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v43
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v42
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v41
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v40
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_or_b32_e32 v0, v39, v0
|
|
; SI-NEXT: s_mov_b32 s6, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v1, v38, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v62, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v37, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v32, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v36, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v61, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v35, v7
|
|
; SI-NEXT: v_or_b32_e32 v8, v60, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v34, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v63, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v33, v11
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v2
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, s6, v3
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v4
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, s6, v5
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v6
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, s6, v7
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v54
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v53
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v52
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v51
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v50
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v49
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v48
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v10
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, s6, v11
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v12
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, s6, v13
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v14
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, s6, v15
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v16
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, s6, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, s6, v19
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, s6, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, s6, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, s6, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 3, v25
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, s6, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 0x30000, v26
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 3, v27
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 0x30000, v27
|
|
; SI-NEXT: .LBB30_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56i16_to_v28f32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_mov_b32_e32 v32, v27
|
|
; VI-NEXT: v_mov_b32_e32 v33, v26
|
|
; VI-NEXT: v_mov_b32_e32 v34, v25
|
|
; VI-NEXT: v_mov_b32_e32 v35, v24
|
|
; VI-NEXT: v_mov_b32_e32 v36, v23
|
|
; VI-NEXT: v_mov_b32_e32 v37, v22
|
|
; VI-NEXT: v_mov_b32_e32 v38, v21
|
|
; VI-NEXT: v_mov_b32_e32 v39, v20
|
|
; VI-NEXT: v_mov_b32_e32 v48, v19
|
|
; VI-NEXT: v_mov_b32_e32 v49, v18
|
|
; VI-NEXT: v_mov_b32_e32 v50, v17
|
|
; VI-NEXT: v_mov_b32_e32 v51, v16
|
|
; VI-NEXT: v_mov_b32_e32 v52, v15
|
|
; VI-NEXT: v_mov_b32_e32 v53, v14
|
|
; VI-NEXT: v_mov_b32_e32 v54, v13
|
|
; VI-NEXT: v_mov_b32_e32 v55, v12
|
|
; VI-NEXT: v_mov_b32_e32 v40, v11
|
|
; VI-NEXT: v_mov_b32_e32 v41, v10
|
|
; VI-NEXT: v_mov_b32_e32 v42, v9
|
|
; VI-NEXT: v_mov_b32_e32 v43, v8
|
|
; VI-NEXT: v_mov_b32_e32 v44, v7
|
|
; VI-NEXT: v_mov_b32_e32 v45, v6
|
|
; VI-NEXT: v_mov_b32_e32 v46, v5
|
|
; VI-NEXT: v_mov_b32_e32 v47, v4
|
|
; VI-NEXT: v_mov_b32_e32 v56, v3
|
|
; VI-NEXT: v_mov_b32_e32 v57, v2
|
|
; VI-NEXT: v_mov_b32_e32 v58, v1
|
|
; VI-NEXT: v_mov_b32_e32 v59, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB30_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_mov_b32_e32 v27, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v27, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v27, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v27, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v27, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v27, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v27, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v27, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v27, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v27, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v27, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v27, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v27, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v27, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v27, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v27, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v27, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v27, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v27, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v27, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v27, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v27, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v24, v27, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v25, v27, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v26, v27, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v27, v27, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v56, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v47, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v46, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v45, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v44, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v43, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v42, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v41, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v40, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v53, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v52, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v51, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v50, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v49, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v39, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v37, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v36, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v35, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v34, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v33, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v32, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: ; implicit-def: $vgpr59
|
|
; VI-NEXT: ; implicit-def: $vgpr58
|
|
; VI-NEXT: ; implicit-def: $vgpr57
|
|
; VI-NEXT: ; implicit-def: $vgpr56
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: .LBB30_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB30_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 3
|
|
; VI-NEXT: v_add_u16_e32 v0, 3, v59
|
|
; VI-NEXT: v_add_u16_sdwa v1, v59, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v58
|
|
; VI-NEXT: v_add_u16_sdwa v3, v58, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v3
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v57
|
|
; VI-NEXT: v_add_u16_sdwa v3, v57, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; VI-NEXT: v_add_u16_e32 v3, 3, v56
|
|
; VI-NEXT: v_add_u16_sdwa v4, v56, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; VI-NEXT: v_add_u16_e32 v4, 3, v47
|
|
; VI-NEXT: v_add_u16_sdwa v5, v47, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; VI-NEXT: v_add_u16_e32 v5, 3, v46
|
|
; VI-NEXT: v_add_u16_sdwa v6, v46, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; VI-NEXT: v_add_u16_e32 v6, 3, v45
|
|
; VI-NEXT: v_add_u16_sdwa v7, v45, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; VI-NEXT: v_add_u16_e32 v7, 3, v44
|
|
; VI-NEXT: v_add_u16_sdwa v8, v44, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; VI-NEXT: v_add_u16_e32 v8, 3, v43
|
|
; VI-NEXT: v_add_u16_sdwa v9, v43, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; VI-NEXT: v_add_u16_e32 v9, 3, v42
|
|
; VI-NEXT: v_add_u16_sdwa v10, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; VI-NEXT: v_add_u16_e32 v10, 3, v41
|
|
; VI-NEXT: v_add_u16_sdwa v11, v41, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v10, v10, v11
|
|
; VI-NEXT: v_add_u16_e32 v11, 3, v40
|
|
; VI-NEXT: v_add_u16_sdwa v12, v40, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; VI-NEXT: v_add_u16_e32 v12, 3, v55
|
|
; VI-NEXT: v_add_u16_sdwa v13, v55, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; VI-NEXT: v_add_u16_e32 v13, 3, v54
|
|
; VI-NEXT: v_add_u16_sdwa v14, v54, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; VI-NEXT: v_add_u16_e32 v14, 3, v53
|
|
; VI-NEXT: v_add_u16_sdwa v15, v53, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; VI-NEXT: v_add_u16_e32 v15, 3, v52
|
|
; VI-NEXT: v_add_u16_sdwa v16, v52, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; VI-NEXT: v_add_u16_e32 v16, 3, v51
|
|
; VI-NEXT: v_add_u16_sdwa v17, v51, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; VI-NEXT: v_add_u16_e32 v17, 3, v50
|
|
; VI-NEXT: v_add_u16_sdwa v18, v50, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; VI-NEXT: v_add_u16_e32 v18, 3, v49
|
|
; VI-NEXT: v_add_u16_sdwa v19, v49, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; VI-NEXT: v_add_u16_e32 v19, 3, v48
|
|
; VI-NEXT: v_add_u16_sdwa v20, v48, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; VI-NEXT: v_add_u16_e32 v20, 3, v39
|
|
; VI-NEXT: v_add_u16_sdwa v21, v39, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; VI-NEXT: v_add_u16_e32 v21, 3, v38
|
|
; VI-NEXT: v_add_u16_sdwa v22, v38, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; VI-NEXT: v_add_u16_e32 v22, 3, v37
|
|
; VI-NEXT: v_add_u16_sdwa v23, v37, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; VI-NEXT: v_add_u16_e32 v23, 3, v36
|
|
; VI-NEXT: v_add_u16_sdwa v24, v36, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; VI-NEXT: v_add_u16_e32 v24, 3, v35
|
|
; VI-NEXT: v_add_u16_sdwa v25, v35, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; VI-NEXT: v_add_u16_e32 v25, 3, v34
|
|
; VI-NEXT: v_add_u16_sdwa v26, v34, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; VI-NEXT: v_add_u16_e32 v26, 3, v33
|
|
; VI-NEXT: v_add_u16_sdwa v28, v33, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; VI-NEXT: v_add_u16_e32 v28, 3, v32
|
|
; VI-NEXT: v_add_u16_sdwa v27, v32, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; VI-NEXT: .LBB30_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56i16_to_v28f32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_mov_b32_e32 v59, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v22
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v13
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v12
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v11
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v10
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v40
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v56, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v57, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v58, v1
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v41
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v47
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v56
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v57
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v58
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v59
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB30_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v37
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: .LBB30_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB30_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB30_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v28f32:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB30_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: .LBB30_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56i16_to_v28f32:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v69, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v71, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v80, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v68, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v67, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v66, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v65, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v64, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v54, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v53, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v52, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v51, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v50, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v49, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v48, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v39, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v38, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v37, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v36, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v35, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v34, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v33, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v32, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v31, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v30, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v29, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB30_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: .LBB30_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <56 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <56 x i16> %a1 to <28 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x i16> %a to <28 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x float> %phi
|
|
}
|
|
|
|
define inreg <28 x float> @bitcast_v56i16_to_v28f32_scalar(<56 x i16> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v56i16_to_v28f32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; SI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; SI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; SI-NEXT: v_writelane_b32 v28, s54, 14
|
|
; SI-NEXT: v_writelane_b32 v28, s55, 15
|
|
; SI-NEXT: v_writelane_b32 v28, s64, 16
|
|
; SI-NEXT: v_writelane_b32 v28, s65, 17
|
|
; SI-NEXT: v_writelane_b32 v28, s66, 18
|
|
; SI-NEXT: v_writelane_b32 v28, s67, 19
|
|
; SI-NEXT: v_writelane_b32 v28, s68, 20
|
|
; SI-NEXT: v_writelane_b32 v28, s69, 21
|
|
; SI-NEXT: v_writelane_b32 v28, s70, 22
|
|
; SI-NEXT: v_writelane_b32 v28, s71, 23
|
|
; SI-NEXT: v_writelane_b32 v28, s80, 24
|
|
; SI-NEXT: v_writelane_b32 v28, s81, 25
|
|
; SI-NEXT: v_writelane_b32 v28, s82, 26
|
|
; SI-NEXT: v_writelane_b32 v28, s83, 27
|
|
; SI-NEXT: v_writelane_b32 v28, s84, 28
|
|
; SI-NEXT: v_writelane_b32 v28, s85, 29
|
|
; SI-NEXT: v_writelane_b32 v28, s86, 30
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s73, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s75, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s77, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s88, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s94, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s31, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s69, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s80, v0
|
|
; SI-NEXT: v_writelane_b32 v28, s87, 31
|
|
; SI-NEXT: s_lshr_b32 s78, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s92, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s95, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s68, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s70, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s81, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s82, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s83, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s84, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s85, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s86, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s87, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s6, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s8, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s10, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s12, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s14, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s72, s73, 16
|
|
; SI-NEXT: s_lshr_b32 s74, s75, 16
|
|
; SI-NEXT: s_lshr_b32 s76, s77, 16
|
|
; SI-NEXT: s_lshr_b32 s79, s88, 16
|
|
; SI-NEXT: s_lshr_b32 s90, s91, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s94, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s31, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s69, 16
|
|
; SI-NEXT: s_lshr_b32 s71, s80, 16
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB31_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s36, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s37, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s38, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s39, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s40, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s41, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s42, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s43, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s44, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s34, 16
|
|
; SI-NEXT: s_or_b32 s45, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s95, 16
|
|
; SI-NEXT: s_or_b32 s46, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s92, 16
|
|
; SI-NEXT: s_or_b32 s47, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s48, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; SI-NEXT: s_or_b32 s49, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s80, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s50, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s51, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s31, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s52, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s94, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s53, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; SI-NEXT: s_or_b32 s54, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s55, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; SI-NEXT: s_or_b32 s56, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; SI-NEXT: s_or_b32 s57, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; SI-NEXT: s_or_b32 s58, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; SI-NEXT: s_or_b32 s59, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s12, 16
|
|
; SI-NEXT: s_or_b32 s60, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s10, 16
|
|
; SI-NEXT: s_or_b32 s61, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s8, 16
|
|
; SI-NEXT: s_or_b32 s62, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s6, 16
|
|
; SI-NEXT: s_or_b32 s63, s4, s5
|
|
; SI-NEXT: s_cbranch_execnz .LBB31_3
|
|
; SI-NEXT: .LBB31_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s17, s17, 3
|
|
; SI-NEXT: s_add_i32 s36, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: s_add_i32 s37, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s19, s19, 3
|
|
; SI-NEXT: s_add_i32 s38, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: s_add_i32 s39, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s21, s21, 3
|
|
; SI-NEXT: s_add_i32 s40, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: s_add_i32 s41, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s42, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s43, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s44, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s34, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s45, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s95, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s46, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s92, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s47, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: s_add_i32 s48, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s80, s80, 3
|
|
; SI-NEXT: s_add_i32 s49, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s80, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s69, s69, 3
|
|
; SI-NEXT: s_add_i32 s50, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s31, s31, 3
|
|
; SI-NEXT: s_add_i32 s51, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s31, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s94, s94, 3
|
|
; SI-NEXT: s_add_i32 s52, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s94, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s91, s91, 3
|
|
; SI-NEXT: s_add_i32 s53, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s88, s88, 3
|
|
; SI-NEXT: s_add_i32 s54, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s77, s77, 3
|
|
; SI-NEXT: s_add_i32 s55, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s75, s75, 3
|
|
; SI-NEXT: s_add_i32 s56, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s73, s73, 3
|
|
; SI-NEXT: s_add_i32 s57, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s15, s15, 3
|
|
; SI-NEXT: s_add_i32 s58, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s13, s13, 3
|
|
; SI-NEXT: s_add_i32 s59, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s12, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s11, s11, 3
|
|
; SI-NEXT: s_add_i32 s60, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s10, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s9, s9, 3
|
|
; SI-NEXT: s_add_i32 s61, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s8, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s7, s7, 3
|
|
; SI-NEXT: s_add_i32 s62, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s6, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s63, s4, 0x30000
|
|
; SI-NEXT: .LBB31_3: ; %end
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_readlane_b32 s87, v28, 31
|
|
; SI-NEXT: v_readlane_b32 s86, v28, 30
|
|
; SI-NEXT: v_readlane_b32 s85, v28, 29
|
|
; SI-NEXT: v_readlane_b32 s84, v28, 28
|
|
; SI-NEXT: v_readlane_b32 s83, v28, 27
|
|
; SI-NEXT: v_readlane_b32 s82, v28, 26
|
|
; SI-NEXT: v_readlane_b32 s81, v28, 25
|
|
; SI-NEXT: v_readlane_b32 s80, v28, 24
|
|
; SI-NEXT: v_readlane_b32 s71, v28, 23
|
|
; SI-NEXT: v_readlane_b32 s70, v28, 22
|
|
; SI-NEXT: v_readlane_b32 s69, v28, 21
|
|
; SI-NEXT: v_readlane_b32 s68, v28, 20
|
|
; SI-NEXT: v_readlane_b32 s67, v28, 19
|
|
; SI-NEXT: v_readlane_b32 s66, v28, 18
|
|
; SI-NEXT: v_readlane_b32 s65, v28, 17
|
|
; SI-NEXT: v_readlane_b32 s64, v28, 16
|
|
; SI-NEXT: v_readlane_b32 s55, v28, 15
|
|
; SI-NEXT: v_readlane_b32 s54, v28, 14
|
|
; SI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB31_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; SI-NEXT: s_branch .LBB31_2
|
|
;
|
|
; VI-LABEL: bitcast_v56i16_to_v28f32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; VI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; VI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; VI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; VI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; VI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; VI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; VI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; VI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; VI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; VI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; VI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; VI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; VI-NEXT: v_writelane_b32 v28, s54, 14
|
|
; VI-NEXT: v_writelane_b32 v28, s55, 15
|
|
; VI-NEXT: v_writelane_b32 v28, s64, 16
|
|
; VI-NEXT: v_writelane_b32 v28, s65, 17
|
|
; VI-NEXT: v_writelane_b32 v28, s66, 18
|
|
; VI-NEXT: v_writelane_b32 v28, s67, 19
|
|
; VI-NEXT: v_writelane_b32 v28, s68, 20
|
|
; VI-NEXT: v_writelane_b32 v28, s69, 21
|
|
; VI-NEXT: v_writelane_b32 v28, s70, 22
|
|
; VI-NEXT: v_writelane_b32 v28, s71, 23
|
|
; VI-NEXT: v_writelane_b32 v28, s80, 24
|
|
; VI-NEXT: v_writelane_b32 v28, s81, 25
|
|
; VI-NEXT: v_writelane_b32 v28, s82, 26
|
|
; VI-NEXT: v_writelane_b32 v28, s83, 27
|
|
; VI-NEXT: v_writelane_b32 v28, s84, 28
|
|
; VI-NEXT: v_writelane_b32 v28, s85, 29
|
|
; VI-NEXT: v_writelane_b32 v28, s86, 30
|
|
; VI-NEXT: v_writelane_b32 v28, s87, 31
|
|
; VI-NEXT: v_readfirstlane_b32 s86, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s87, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s73, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s77, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s88, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s34, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s69, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s81, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s84, v0
|
|
; VI-NEXT: s_lshr_b32 s79, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s67, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s70, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s80, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s82, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s85, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s7, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s11, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s13, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s15, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s10, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s64, s86, 16
|
|
; VI-NEXT: s_lshr_b32 s65, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s66, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s14, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s87, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s73, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s77, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s88, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s91, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s34, 16
|
|
; VI-NEXT: s_lshr_b32 s68, s69, 16
|
|
; VI-NEXT: s_lshr_b32 s71, s81, 16
|
|
; VI-NEXT: s_lshr_b32 s83, s84, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB31_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_or_b32 s36, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_or_b32 s38, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s10, 16
|
|
; VI-NEXT: s_or_b32 s39, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; VI-NEXT: s_or_b32 s37, s40, s41
|
|
; VI-NEXT: s_or_b32 s40, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_or_b32 s41, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s42, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s43, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; VI-NEXT: s_or_b32 s44, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s45, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s5, s67, 16
|
|
; VI-NEXT: s_or_b32 s46, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s47, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s48, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s49, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s84
|
|
; VI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; VI-NEXT: s_or_b32 s50, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s81
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s51, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s69
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s52, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s34
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s53, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s91
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s54, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s88
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s55, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s77
|
|
; VI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; VI-NEXT: s_or_b32 s56, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s73
|
|
; VI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; VI-NEXT: s_or_b32 s57, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; VI-NEXT: s_or_b32 s58, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s87
|
|
; VI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; VI-NEXT: s_or_b32 s59, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; VI-NEXT: s_or_b32 s60, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s5, s66, 16
|
|
; VI-NEXT: s_or_b32 s61, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s5, s65, 16
|
|
; VI-NEXT: s_or_b32 s62, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s86
|
|
; VI-NEXT: s_lshl_b32 s5, s64, 16
|
|
; VI-NEXT: s_or_b32 s63, s4, s5
|
|
; VI-NEXT: s_cbranch_execnz .LBB31_3
|
|
; VI-NEXT: .LBB31_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_and_b32 s5, s17, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s10, s10, 16
|
|
; VI-NEXT: s_or_b32 s5, s10, s5
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s36, s4, 0x30000
|
|
; VI-NEXT: s_add_i32 s37, s5, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s38, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s39, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s40, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s41, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s42, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s43, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s44, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s45, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s67, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s46, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s47, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s48, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s84, s84, 3
|
|
; VI-NEXT: s_add_i32 s49, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s84, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s81, s81, 3
|
|
; VI-NEXT: s_add_i32 s50, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s81, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s69, s69, 3
|
|
; VI-NEXT: s_add_i32 s51, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s34, s34, 3
|
|
; VI-NEXT: s_add_i32 s52, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s34, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s91, s91, 3
|
|
; VI-NEXT: s_add_i32 s53, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s88, s88, 3
|
|
; VI-NEXT: s_add_i32 s54, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s77, s77, 3
|
|
; VI-NEXT: s_add_i32 s55, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s75, s73, 3
|
|
; VI-NEXT: s_add_i32 s56, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s73, s12, 3
|
|
; VI-NEXT: s_add_i32 s57, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s15, s87, 3
|
|
; VI-NEXT: s_add_i32 s58, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s13, s8, 3
|
|
; VI-NEXT: s_add_i32 s59, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s11, s9, 3
|
|
; VI-NEXT: s_add_i32 s60, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s66, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s9, s6, 3
|
|
; VI-NEXT: s_add_i32 s61, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s65, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s7, s86, 3
|
|
; VI-NEXT: s_add_i32 s62, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s64, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s63, s4, 0x30000
|
|
; VI-NEXT: .LBB31_3: ; %end
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_readlane_b32 s87, v28, 31
|
|
; VI-NEXT: v_readlane_b32 s86, v28, 30
|
|
; VI-NEXT: v_readlane_b32 s85, v28, 29
|
|
; VI-NEXT: v_readlane_b32 s84, v28, 28
|
|
; VI-NEXT: v_readlane_b32 s83, v28, 27
|
|
; VI-NEXT: v_readlane_b32 s82, v28, 26
|
|
; VI-NEXT: v_readlane_b32 s81, v28, 25
|
|
; VI-NEXT: v_readlane_b32 s80, v28, 24
|
|
; VI-NEXT: v_readlane_b32 s71, v28, 23
|
|
; VI-NEXT: v_readlane_b32 s70, v28, 22
|
|
; VI-NEXT: v_readlane_b32 s69, v28, 21
|
|
; VI-NEXT: v_readlane_b32 s68, v28, 20
|
|
; VI-NEXT: v_readlane_b32 s67, v28, 19
|
|
; VI-NEXT: v_readlane_b32 s66, v28, 18
|
|
; VI-NEXT: v_readlane_b32 s65, v28, 17
|
|
; VI-NEXT: v_readlane_b32 s64, v28, 16
|
|
; VI-NEXT: v_readlane_b32 s55, v28, 15
|
|
; VI-NEXT: v_readlane_b32 s54, v28, 14
|
|
; VI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; VI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; VI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; VI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; VI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; VI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; VI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; VI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; VI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; VI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; VI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB31_4:
|
|
; VI-NEXT: ; implicit-def: $vgpr29 : SGPR spill to VGPR lane
|
|
; VI-NEXT: s_mov_b32 vcc_lo, s64
|
|
; VI-NEXT: v_writelane_b32 v29, s75, 0
|
|
; VI-NEXT: s_mov_b32 vcc_hi, s65
|
|
; VI-NEXT: s_mov_b32 s75, s73
|
|
; VI-NEXT: s_mov_b32 s73, s12
|
|
; VI-NEXT: v_writelane_b32 v29, s25, 1
|
|
; VI-NEXT: s_mov_b32 s25, s91
|
|
; VI-NEXT: s_mov_b32 s91, s26
|
|
; VI-NEXT: s_mov_b32 s26, s27
|
|
; VI-NEXT: s_mov_b32 s27, s88
|
|
; VI-NEXT: s_mov_b32 s88, s28
|
|
; VI-NEXT: s_mov_b32 s28, s77
|
|
; VI-NEXT: s_mov_b32 s77, s29
|
|
; VI-NEXT: s_mov_b32 s29, s15
|
|
; VI-NEXT: s_mov_b32 s15, s87
|
|
; VI-NEXT: s_mov_b32 s87, s85
|
|
; VI-NEXT: s_mov_b32 s85, s82
|
|
; VI-NEXT: s_mov_b32 s82, s71
|
|
; VI-NEXT: s_mov_b32 s71, s68
|
|
; VI-NEXT: s_mov_b32 s68, s35
|
|
; VI-NEXT: s_mov_b32 s35, s31
|
|
; VI-NEXT: s_mov_b32 s31, s30
|
|
; VI-NEXT: s_mov_b32 s30, s90
|
|
; VI-NEXT: s_mov_b32 s90, s89
|
|
; VI-NEXT: s_mov_b32 s89, s79
|
|
; VI-NEXT: s_mov_b32 s79, s78
|
|
; VI-NEXT: s_mov_b32 s78, s76
|
|
; VI-NEXT: s_mov_b32 s76, s74
|
|
; VI-NEXT: s_mov_b32 s74, s72
|
|
; VI-NEXT: s_mov_b32 s72, s14
|
|
; VI-NEXT: s_mov_b32 s14, s66
|
|
; VI-NEXT: s_mov_b32 s12, s10
|
|
; VI-NEXT: s_mov_b32 s10, s13
|
|
; VI-NEXT: s_mov_b32 s13, s8
|
|
; VI-NEXT: s_mov_b32 s8, s11
|
|
; VI-NEXT: s_mov_b32 s11, s9
|
|
; VI-NEXT: s_mov_b32 s9, s6
|
|
; VI-NEXT: s_mov_b32 s6, s7
|
|
; VI-NEXT: s_mov_b32 s7, s86
|
|
; VI-NEXT: s_mov_b32 s86, s83
|
|
; VI-NEXT: s_mov_b32 s83, s80
|
|
; VI-NEXT: s_mov_b32 s80, s70
|
|
; VI-NEXT: s_mov_b32 s70, s67
|
|
; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; VI-NEXT: s_mov_b32 s67, s70
|
|
; VI-NEXT: s_mov_b32 s70, s80
|
|
; VI-NEXT: s_mov_b32 s80, s83
|
|
; VI-NEXT: s_mov_b32 s83, s86
|
|
; VI-NEXT: s_mov_b32 s86, s7
|
|
; VI-NEXT: s_mov_b32 s7, s6
|
|
; VI-NEXT: s_mov_b32 s6, s9
|
|
; VI-NEXT: s_mov_b32 s9, s11
|
|
; VI-NEXT: s_mov_b32 s11, s8
|
|
; VI-NEXT: s_mov_b32 s8, s13
|
|
; VI-NEXT: s_mov_b32 s13, s10
|
|
; VI-NEXT: s_mov_b32 s10, s12
|
|
; VI-NEXT: s_mov_b32 s66, s14
|
|
; VI-NEXT: s_mov_b32 s14, s72
|
|
; VI-NEXT: s_mov_b32 s72, s74
|
|
; VI-NEXT: s_mov_b32 s74, s76
|
|
; VI-NEXT: s_mov_b32 s76, s78
|
|
; VI-NEXT: s_mov_b32 s78, s79
|
|
; VI-NEXT: s_mov_b32 s79, s89
|
|
; VI-NEXT: s_mov_b32 s89, s90
|
|
; VI-NEXT: s_mov_b32 s90, s30
|
|
; VI-NEXT: s_mov_b32 s30, s31
|
|
; VI-NEXT: s_mov_b32 s31, s35
|
|
; VI-NEXT: s_mov_b32 s35, s68
|
|
; VI-NEXT: s_mov_b32 s68, s71
|
|
; VI-NEXT: s_mov_b32 s71, s82
|
|
; VI-NEXT: s_mov_b32 s82, s85
|
|
; VI-NEXT: s_mov_b32 s85, s87
|
|
; VI-NEXT: s_mov_b32 s87, s15
|
|
; VI-NEXT: s_mov_b32 s15, s29
|
|
; VI-NEXT: s_mov_b32 s29, s77
|
|
; VI-NEXT: s_mov_b32 s77, s28
|
|
; VI-NEXT: s_mov_b32 s28, s88
|
|
; VI-NEXT: s_mov_b32 s88, s27
|
|
; VI-NEXT: s_mov_b32 s27, s26
|
|
; VI-NEXT: s_mov_b32 s26, s91
|
|
; VI-NEXT: s_mov_b32 s91, s25
|
|
; VI-NEXT: v_readlane_b32 s25, v29, 1
|
|
; VI-NEXT: s_mov_b32 s12, s73
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: v_readlane_b32 s75, v29, 0
|
|
; VI-NEXT: s_mov_b32 s65, vcc_hi
|
|
; VI-NEXT: s_mov_b32 s64, vcc_lo
|
|
; VI-NEXT: s_branch .LBB31_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v56i16_to_v28f32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: s_lshr_b32 s12, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s13, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s14, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s15, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s88, v5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v3
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v2
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: s_lshr_b32 s4, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s5, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s6, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s7, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s8, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s9, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s10, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s63, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s62, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s61, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s60, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s59, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s58, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s57, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s56, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s88, 16
|
|
; GFX9-NEXT: s_lshr_b32 s17, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s18, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s19, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s20, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s21, s12, 16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_cmp_lg_u32 s42, 0
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s50, s12, s21
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s51, s13, s20
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s52, s14, s19
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s53, s15, s18
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s54, s16, s17
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s55, s88, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s56, s56, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s57, s57, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s58, s58, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s59, s59, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s60, s60, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s61, s61, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s62, s62, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s63, s63, s72
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB31_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB31_4
|
|
; GFX9-NEXT: .LBB31_2: ; %cmp.true
|
|
; GFX9-NEXT: v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_branch .LBB31_5
|
|
; GFX9-NEXT: .LBB31_3:
|
|
; GFX9-NEXT: s_branch .LBB31_2
|
|
; GFX9-NEXT: .LBB31_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB31_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v56i16_to_v28f32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s44, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s74, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s76, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s77, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s78, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s88, v10
|
|
; GFX11-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s0, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s44, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s56, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s58, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s59, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s60, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s62, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s74, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s76, 16
|
|
; GFX11-NEXT: s_lshr_b32 s90, s77, 16
|
|
; GFX11-NEXT: s_lshr_b32 s91, s78, 16
|
|
; GFX11-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s88, 0
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s78, s91
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s77, s90
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s76, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s74, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s62, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s60, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s59, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s58, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s56, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s44, s57
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB31_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB31_4
|
|
; GFX11-NEXT: .LBB31_2: ; %cmp.true
|
|
; GFX11-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v18, s18, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v19, s19, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v20, s20, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v21, s21, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v22, s22, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v23, s23, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v24, s24, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v25, s25, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v26, s26, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v27, s27, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB31_3:
|
|
; GFX11-NEXT: s_branch .LBB31_2
|
|
; GFX11-NEXT: .LBB31_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <56 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <56 x i16> %a1 to <28 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x i16> %a to <28 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x float> %phi
|
|
}
|
|
|
|
define <56 x half> @bitcast_v28f32_to_v56f16(<28 x float> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v28f32_to_v56f16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB32_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB32_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB32_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; SI-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; SI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; SI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; SI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; SI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; SI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; SI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; SI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; SI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; SI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; SI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; SI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; SI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; SI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; SI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; SI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; SI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; SI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; SI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; SI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; SI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; SI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; SI-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; SI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; SI-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; SI-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB32_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v47
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v40
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v48
|
|
; SI-NEXT: v_lshlrev_b32_e32 v48, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v41
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v53
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v51
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v48
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v38
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28f32_to_v56f16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB32_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB32_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB32_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; VI-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; VI-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; VI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; VI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; VI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; VI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; VI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; VI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; VI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; VI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; VI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; VI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; VI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; VI-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB32_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28f32_to_v56f16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB32_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB32_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB32_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f32_e32 v27, 1.0, v27
|
|
; GFX9-NEXT: v_add_f32_e32 v26, 1.0, v26
|
|
; GFX9-NEXT: v_add_f32_e32 v25, 1.0, v25
|
|
; GFX9-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; GFX9-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; GFX9-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; GFX9-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; GFX9-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; GFX9-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; GFX9-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; GFX9-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; GFX9-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; GFX9-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; GFX9-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB32_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v47, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v46, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v45, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v44, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v43, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v42, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v41, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v40, v7, s4
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v8, v55, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v54, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v53, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v52, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v51, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v49, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v48, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v39, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v38, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v37, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v36, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v35, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v34, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v33, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v32, v23, s4
|
|
; GFX9-NEXT: v_perm_b32 v24, v31, v24, s4
|
|
; GFX9-NEXT: v_perm_b32 v25, v30, v25, s4
|
|
; GFX9-NEXT: v_perm_b32 v26, v29, v26, s4
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s4
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v28f32_to_v56f16:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB32_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
|
|
; GFX11-TRUE16-NEXT: .LBB32_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v28f32_to_v56f16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr67
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB32_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB32_2: ; %Flow
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB32_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB32_4: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v71, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v69, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v68, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v67, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v66, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v65, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v64, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v55, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v54, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v53, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v52, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v51, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v49, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v39, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v37, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v36, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v35, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v34, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v33, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v32, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v31, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v30, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v29, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <28 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <28 x float> %a1 to <56 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x float> %a to <56 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x half> %phi
|
|
}
|
|
|
|
define inreg <56 x half> @bitcast_v28f32_to_v56f16_scalar(<28 x float> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v28f32_to_v56f16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v58, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v58, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v58, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v58, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v58, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v58, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v58, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v58, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v58, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v58, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v58, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v58, s51, 11
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v14
|
|
; SI-NEXT: v_writelane_b32 v58, s52, 12
|
|
; SI-NEXT: v_readfirstlane_b32 s5, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s40, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v0
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_writelane_b32 v58, s53, 13
|
|
; SI-NEXT: s_cbranch_scc0 .LBB33_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s53, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB33_4
|
|
; SI-NEXT: .LBB33_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e64 v27, s5, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v26, s4, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v25, s7, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v24, s6, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[28:29], v[26:27], 16
|
|
; SI-NEXT: v_add_f32_e64 v23, s9, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v22, s8, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[29:30], v[24:25], 16
|
|
; SI-NEXT: v_add_f32_e64 v21, s11, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v20, s10, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[30:31], v[22:23], 16
|
|
; SI-NEXT: v_add_f32_e64 v19, s13, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v18, s12, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[31:32], v[20:21], 16
|
|
; SI-NEXT: v_add_f32_e64 v17, s15, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v16, s14, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[32:33], v[18:19], 16
|
|
; SI-NEXT: v_add_f32_e64 v15, s41, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v14, s40, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[33:34], v[16:17], 16
|
|
; SI-NEXT: v_add_f32_e64 v13, s29, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v12, s28, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[34:35], v[14:15], 16
|
|
; SI-NEXT: v_add_f32_e64 v5, s21, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v4, s20, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v11, s27, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v10, s26, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[35:36], v[12:13], 16
|
|
; SI-NEXT: v_add_f32_e64 v3, s19, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v2, s18, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v9, s25, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v8, s24, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[36:37], v[10:11], 16
|
|
; SI-NEXT: v_lshr_b64 v[48:49], v[4:5], 16
|
|
; SI-NEXT: v_add_f32_e64 v1, s17, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v0, s16, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v7, s23, 1.0
|
|
; SI-NEXT: v_add_f32_e64 v6, s22, 1.0
|
|
; SI-NEXT: v_lshr_b64 v[37:38], v[8:9], 16
|
|
; SI-NEXT: v_lshr_b64 v[49:50], v[2:3], 16
|
|
; SI-NEXT: v_lshr_b64 v[38:39], v[6:7], 16
|
|
; SI-NEXT: v_lshr_b64 v[50:51], v[0:1], 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v1
|
|
; SI-NEXT: s_branch .LBB33_5
|
|
; SI-NEXT: .LBB33_3:
|
|
; SI-NEXT: ; implicit-def: $sgpr92
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: ; implicit-def: $sgpr31
|
|
; SI-NEXT: ; implicit-def: $sgpr88
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr78
|
|
; SI-NEXT: ; implicit-def: $sgpr35
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr36
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr37
|
|
; SI-NEXT: ; implicit-def: $sgpr72
|
|
; SI-NEXT: ; implicit-def: $sgpr38
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr39
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr48
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr49
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr50
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr51
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr52
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr53
|
|
; SI-NEXT: s_branch .LBB33_2
|
|
; SI-NEXT: .LBB33_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s40
|
|
; SI-NEXT: v_mov_b32_e32 v15, s41
|
|
; SI-NEXT: v_mov_b32_e32 v16, s14
|
|
; SI-NEXT: v_mov_b32_e32 v17, s15
|
|
; SI-NEXT: v_mov_b32_e32 v18, s12
|
|
; SI-NEXT: v_mov_b32_e32 v19, s13
|
|
; SI-NEXT: v_mov_b32_e32 v20, s10
|
|
; SI-NEXT: v_mov_b32_e32 v21, s11
|
|
; SI-NEXT: v_mov_b32_e32 v22, s8
|
|
; SI-NEXT: v_mov_b32_e32 v23, s9
|
|
; SI-NEXT: v_mov_b32_e32 v24, s6
|
|
; SI-NEXT: v_mov_b32_e32 v25, s7
|
|
; SI-NEXT: v_mov_b32_e32 v26, s4
|
|
; SI-NEXT: v_mov_b32_e32 v27, s5
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v57, s30
|
|
; SI-NEXT: v_mov_b32_e32 v56, s31
|
|
; SI-NEXT: v_mov_b32_e32 v47, s34
|
|
; SI-NEXT: v_mov_b32_e32 v46, s35
|
|
; SI-NEXT: v_mov_b32_e32 v45, s36
|
|
; SI-NEXT: v_mov_b32_e32 v44, s37
|
|
; SI-NEXT: v_mov_b32_e32 v43, s38
|
|
; SI-NEXT: v_mov_b32_e32 v42, s39
|
|
; SI-NEXT: v_mov_b32_e32 v41, s48
|
|
; SI-NEXT: v_mov_b32_e32 v40, s49
|
|
; SI-NEXT: v_mov_b32_e32 v55, s50
|
|
; SI-NEXT: v_mov_b32_e32 v54, s51
|
|
; SI-NEXT: v_mov_b32_e32 v53, s52
|
|
; SI-NEXT: v_mov_b32_e32 v52, s53
|
|
; SI-NEXT: v_mov_b32_e32 v28, s42
|
|
; SI-NEXT: v_mov_b32_e32 v29, s44
|
|
; SI-NEXT: v_mov_b32_e32 v30, s46
|
|
; SI-NEXT: v_mov_b32_e32 v31, s56
|
|
; SI-NEXT: v_mov_b32_e32 v32, s58
|
|
; SI-NEXT: v_mov_b32_e32 v33, s60
|
|
; SI-NEXT: v_mov_b32_e32 v34, s62
|
|
; SI-NEXT: v_mov_b32_e32 v35, s72
|
|
; SI-NEXT: v_mov_b32_e32 v36, s74
|
|
; SI-NEXT: v_mov_b32_e32 v37, s76
|
|
; SI-NEXT: v_mov_b32_e32 v38, s78
|
|
; SI-NEXT: v_mov_b32_e32 v48, s88
|
|
; SI-NEXT: v_mov_b32_e32 v49, s90
|
|
; SI-NEXT: v_mov_b32_e32 v50, s92
|
|
; SI-NEXT: .LBB33_5: ; %end
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v50
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v39
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v57
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v49
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v39
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v56
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v48
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v41
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v40
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v39
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v38
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v37
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: v_readlane_b32 s53, v58, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v58, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v58, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v58, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v58, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v58, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v58, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v58, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v58, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v58, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v58, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v58, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v58, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v58, 0
|
|
; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v28f32_to_v56f16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v56, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v56, s31, 1
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v56, s34, 2
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_writelane_b32 v56, s35, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB33_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s44, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s42, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB33_4
|
|
; VI-NEXT: .LBB33_2: ; %cmp.true
|
|
; VI-NEXT: v_add_f32_e64 v27, s6, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v26, s7, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v25, s8, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v24, s9, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v23, s10, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v22, s11, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v21, s12, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v20, s13, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v19, s14, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v18, s15, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v17, s40, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v16, s41, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v15, s42, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v14, s43, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v13, s29, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v12, s28, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v11, s27, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v10, s26, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v9, s25, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v8, s24, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v7, s23, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v6, s22, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v5, s21, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v4, s20, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v3, s19, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v2, s18, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v1, s17, 1.0
|
|
; VI-NEXT: v_add_f32_e64 v0, s16, 1.0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: s_branch .LBB33_5
|
|
; VI-NEXT: .LBB33_3:
|
|
; VI-NEXT: ; implicit-def: $sgpr35
|
|
; VI-NEXT: ; implicit-def: $sgpr34
|
|
; VI-NEXT: ; implicit-def: $sgpr31
|
|
; VI-NEXT: ; implicit-def: $sgpr30
|
|
; VI-NEXT: ; implicit-def: $sgpr91
|
|
; VI-NEXT: ; implicit-def: $sgpr90
|
|
; VI-NEXT: ; implicit-def: $sgpr89
|
|
; VI-NEXT: ; implicit-def: $sgpr88
|
|
; VI-NEXT: ; implicit-def: $sgpr79
|
|
; VI-NEXT: ; implicit-def: $sgpr78
|
|
; VI-NEXT: ; implicit-def: $sgpr77
|
|
; VI-NEXT: ; implicit-def: $sgpr76
|
|
; VI-NEXT: ; implicit-def: $sgpr75
|
|
; VI-NEXT: ; implicit-def: $sgpr74
|
|
; VI-NEXT: ; implicit-def: $sgpr73
|
|
; VI-NEXT: ; implicit-def: $sgpr72
|
|
; VI-NEXT: ; implicit-def: $sgpr63
|
|
; VI-NEXT: ; implicit-def: $sgpr62
|
|
; VI-NEXT: ; implicit-def: $sgpr61
|
|
; VI-NEXT: ; implicit-def: $sgpr60
|
|
; VI-NEXT: ; implicit-def: $sgpr59
|
|
; VI-NEXT: ; implicit-def: $sgpr58
|
|
; VI-NEXT: ; implicit-def: $sgpr57
|
|
; VI-NEXT: ; implicit-def: $sgpr56
|
|
; VI-NEXT: ; implicit-def: $sgpr47
|
|
; VI-NEXT: ; implicit-def: $sgpr46
|
|
; VI-NEXT: ; implicit-def: $sgpr45
|
|
; VI-NEXT: ; implicit-def: $sgpr44
|
|
; VI-NEXT: s_branch .LBB33_2
|
|
; VI-NEXT: .LBB33_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v14, s43
|
|
; VI-NEXT: v_mov_b32_e32 v15, s42
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: v_mov_b32_e32 v47, s35
|
|
; VI-NEXT: v_mov_b32_e32 v46, s34
|
|
; VI-NEXT: v_mov_b32_e32 v45, s31
|
|
; VI-NEXT: v_mov_b32_e32 v44, s30
|
|
; VI-NEXT: v_mov_b32_e32 v43, s91
|
|
; VI-NEXT: v_mov_b32_e32 v42, s90
|
|
; VI-NEXT: v_mov_b32_e32 v41, s89
|
|
; VI-NEXT: v_mov_b32_e32 v40, s88
|
|
; VI-NEXT: v_mov_b32_e32 v55, s79
|
|
; VI-NEXT: v_mov_b32_e32 v54, s78
|
|
; VI-NEXT: v_mov_b32_e32 v53, s77
|
|
; VI-NEXT: v_mov_b32_e32 v52, s76
|
|
; VI-NEXT: v_mov_b32_e32 v51, s75
|
|
; VI-NEXT: v_mov_b32_e32 v50, s74
|
|
; VI-NEXT: v_mov_b32_e32 v49, s73
|
|
; VI-NEXT: v_mov_b32_e32 v48, s72
|
|
; VI-NEXT: v_mov_b32_e32 v39, s63
|
|
; VI-NEXT: v_mov_b32_e32 v38, s62
|
|
; VI-NEXT: v_mov_b32_e32 v37, s61
|
|
; VI-NEXT: v_mov_b32_e32 v36, s60
|
|
; VI-NEXT: v_mov_b32_e32 v35, s59
|
|
; VI-NEXT: v_mov_b32_e32 v34, s58
|
|
; VI-NEXT: v_mov_b32_e32 v33, s57
|
|
; VI-NEXT: v_mov_b32_e32 v32, s56
|
|
; VI-NEXT: v_mov_b32_e32 v31, s47
|
|
; VI-NEXT: v_mov_b32_e32 v30, s46
|
|
; VI-NEXT: v_mov_b32_e32 v29, s45
|
|
; VI-NEXT: v_mov_b32_e32 v28, s44
|
|
; VI-NEXT: .LBB33_5: ; %end
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_readlane_b32 s35, v56, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v56, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v56, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v56, 0
|
|
; VI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v28f32_to_v56f16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB33_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s44, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s42, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s43, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB33_4
|
|
; GFX9-NEXT: .LBB33_2: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f32_e64 v27, s6, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v26, s7, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v25, s8, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v24, s9, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v23, s10, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v22, s11, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v21, s12, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v20, s13, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v19, s14, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v18, s15, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v17, s40, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v16, s41, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v15, s42, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v14, s43, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v13, s29, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v12, s28, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v11, s27, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v10, s26, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v9, s25, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v8, s24, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v7, s23, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v6, s22, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v5, s21, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v4, s20, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v3, s19, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v2, s18, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v1, s17, 1.0
|
|
; GFX9-NEXT: v_add_f32_e64 v0, s16, 1.0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: s_branch .LBB33_5
|
|
; GFX9-NEXT: .LBB33_3:
|
|
; GFX9-NEXT: ; implicit-def: $sgpr95
|
|
; GFX9-NEXT: ; implicit-def: $sgpr94
|
|
; GFX9-NEXT: ; implicit-def: $sgpr93
|
|
; GFX9-NEXT: ; implicit-def: $sgpr92
|
|
; GFX9-NEXT: ; implicit-def: $sgpr91
|
|
; GFX9-NEXT: ; implicit-def: $sgpr90
|
|
; GFX9-NEXT: ; implicit-def: $sgpr89
|
|
; GFX9-NEXT: ; implicit-def: $sgpr88
|
|
; GFX9-NEXT: ; implicit-def: $sgpr79
|
|
; GFX9-NEXT: ; implicit-def: $sgpr78
|
|
; GFX9-NEXT: ; implicit-def: $sgpr77
|
|
; GFX9-NEXT: ; implicit-def: $sgpr76
|
|
; GFX9-NEXT: ; implicit-def: $sgpr75
|
|
; GFX9-NEXT: ; implicit-def: $sgpr74
|
|
; GFX9-NEXT: ; implicit-def: $sgpr73
|
|
; GFX9-NEXT: ; implicit-def: $sgpr72
|
|
; GFX9-NEXT: ; implicit-def: $sgpr63
|
|
; GFX9-NEXT: ; implicit-def: $sgpr62
|
|
; GFX9-NEXT: ; implicit-def: $sgpr61
|
|
; GFX9-NEXT: ; implicit-def: $sgpr60
|
|
; GFX9-NEXT: ; implicit-def: $sgpr59
|
|
; GFX9-NEXT: ; implicit-def: $sgpr58
|
|
; GFX9-NEXT: ; implicit-def: $sgpr57
|
|
; GFX9-NEXT: ; implicit-def: $sgpr56
|
|
; GFX9-NEXT: ; implicit-def: $sgpr47
|
|
; GFX9-NEXT: ; implicit-def: $sgpr46
|
|
; GFX9-NEXT: ; implicit-def: $sgpr45
|
|
; GFX9-NEXT: ; implicit-def: $sgpr44
|
|
; GFX9-NEXT: s_branch .LBB33_2
|
|
; GFX9-NEXT: .LBB33_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, s95
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, s94
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, s93
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, s92
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, s91
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, s90
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, s89
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, s88
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, s79
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, s78
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, s77
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, s76
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, s75
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, s74
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, s73
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, s72
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s44
|
|
; GFX9-NEXT: .LBB33_5: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX9-NEXT: v_lshl_or_b32 v0, v47, 16, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v1, v46, 16, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v2, v45, 16, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v3, v44, 16, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v4, v43, 16, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v5, v42, 16, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v6, v41, 16, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v7, v40, 16, v7
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX9-NEXT: v_lshl_or_b32 v8, v55, 16, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v9, v54, 16, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v53, 16, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v52, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v51, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v50, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v49, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v48, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v39, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v38, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v37, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v36, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v35, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v34, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v33, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v32, 16, v23
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v31, 16, v24
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v30, 16, v25
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v29, 16, v26
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v28, 16, v27
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v28f32_to_v56f16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB33_3
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s4, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s5, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s6, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s7, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s8, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s9, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s10, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s12, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s56, s13, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s57, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s58, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s59, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s60, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s61, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s62, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s63, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s72, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s73, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s74, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s75, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s76, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s77, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s78, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s79, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s88, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s89, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s90, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB33_4
|
|
; GFX11-TRUE16-NEXT: .LBB33_2: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v27, s4, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v26, s5, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v25, s6, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v24, s7, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v23, s8, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v22, s9, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v21, s10, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v20, s11, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v19, s12, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v18, s13, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v17, s29, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v16, s28, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v15, s27, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v14, s26, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v13, s25, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v12, s24, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v11, s23, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v10, s22, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v9, s21, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v8, s20, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v7, s19, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v6, s18, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v5, s17, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v4, s16, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v3, s3, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v2, s2, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v1, s1, 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f32_e64 v0, s0, 1.0
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB33_5
|
|
; GFX11-TRUE16-NEXT: .LBB33_3:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr90
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB33_2
|
|
; GFX11-TRUE16-NEXT: .LBB33_4:
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v71, s90 :: v_dual_mov_b32 v70, s89
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v69, s88 :: v_dual_mov_b32 v68, s79
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v67, s78 :: v_dual_mov_b32 v66, s77
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v65, s76 :: v_dual_mov_b32 v64, s75
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v55, s74 :: v_dual_mov_b32 v54, s73
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, s72 :: v_dual_mov_b32 v52, s63
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v51, s62 :: v_dual_mov_b32 v50, s61
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v49, s60 :: v_dual_mov_b32 v48, s59
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, s58 :: v_dual_mov_b32 v38, s57
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, s56 :: v_dual_mov_b32 v36, s47
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15
|
|
; GFX11-TRUE16-NEXT: .LBB33_5: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v71.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v70.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v69.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v68.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v67.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v66.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v65.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v64.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v31.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v28.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v28f32_to_v56f16_scalar:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB33_3
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s4, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s5, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s6, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s7, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s8, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s9, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s10, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s12, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s56, s13, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s57, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s58, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s59, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s60, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s61, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s62, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s63, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s72, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s73, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s74, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s75, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s76, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s77, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s78, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s79, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s88, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s89, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s90, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB33_4
|
|
; GFX11-FAKE16-NEXT: .LBB33_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v23, s4, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v24, s5, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v25, s6, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v26, s7, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v27, s8, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v18, s9, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v19, s10, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v20, s11, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v21, s12, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v22, s13, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v13, s29, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v14, s28, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v15, s27, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v16, s26, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v17, s25, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v8, s24, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v9, s23, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v10, s22, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v11, s21, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v12, s20, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v3, s19, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v4, s18, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v5, s17, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v6, s16, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v7, s3, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v0, s2, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v1, s1, 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f32_e64 v2, s0, 1.0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB33_5
|
|
; GFX11-FAKE16-NEXT: .LBB33_3:
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr90
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB33_2
|
|
; GFX11-FAKE16-NEXT: .LBB33_4:
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v7, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s18 :: v_dual_mov_b32 v3, s19
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s20 :: v_dual_mov_b32 v11, s21
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v9, s23
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s24 :: v_dual_mov_b32 v17, s25
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s28 :: v_dual_mov_b32 v13, s29
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v21, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v19, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v18, s9 :: v_dual_mov_b32 v27, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v26, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s5 :: v_dual_mov_b32 v23, s4
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v71, s90 :: v_dual_mov_b32 v70, s89
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v69, s88 :: v_dual_mov_b32 v68, s79
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v67, s78 :: v_dual_mov_b32 v66, s77
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v65, s76 :: v_dual_mov_b32 v64, s75
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v55, s74 :: v_dual_mov_b32 v54, s73
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v53, s72 :: v_dual_mov_b32 v52, s63
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v51, s62 :: v_dual_mov_b32 v50, s61
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v49, s60 :: v_dual_mov_b32 v48, s59
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v39, s58 :: v_dual_mov_b32 v38, s57
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, s56 :: v_dual_mov_b32 v36, s47
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15
|
|
; GFX11-FAKE16-NEXT: .LBB33_5: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v80, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v69, 16, v80
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v69, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v70, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v68, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v67, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v65, 16, v69
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v64, 16, v70
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v65, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v55, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v54, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v52, 16, v64
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v51, 16, v65
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v13
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v50, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v49, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v39, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v38, 16, v52
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v19
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v23
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v31, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v29, 16, v33
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v28, 16, v34
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <28 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <28 x float> %a1 to <56 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <28 x float> %a to <56 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x half> %phi
|
|
}
|
|
|
|
define <28 x float> @bitcast_v56f16_to_v28f32(<56 x half> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v56f16_to_v28f32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_mov_b32_e32 v59, v0
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v22
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v21
|
|
; SI-NEXT: v_mov_b32_e32 v48, v19
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v20
|
|
; SI-NEXT: v_mov_b32_e32 v49, v18
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; SI-NEXT: v_mov_b32_e32 v50, v17
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; SI-NEXT: v_mov_b32_e32 v51, v16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; SI-NEXT: v_mov_b32_e32 v52, v15
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; SI-NEXT: v_mov_b32_e32 v53, v14
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; SI-NEXT: v_mov_b32_e32 v54, v13
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; SI-NEXT: v_mov_b32_e32 v55, v12
|
|
; SI-NEXT: v_mov_b32_e32 v40, v11
|
|
; SI-NEXT: v_mov_b32_e32 v41, v10
|
|
; SI-NEXT: v_mov_b32_e32 v42, v9
|
|
; SI-NEXT: v_mov_b32_e32 v43, v8
|
|
; SI-NEXT: v_mov_b32_e32 v44, v7
|
|
; SI-NEXT: v_mov_b32_e32 v45, v6
|
|
; SI-NEXT: v_mov_b32_e32 v46, v5
|
|
; SI-NEXT: v_mov_b32_e32 v47, v4
|
|
; SI-NEXT: v_mov_b32_e32 v56, v3
|
|
; SI-NEXT: v_mov_b32_e32 v57, v2
|
|
; SI-NEXT: v_mov_b32_e32 v58, v1
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v40
|
|
; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v41
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v42
|
|
; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v43
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v44
|
|
; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v46
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v47
|
|
; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v56
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v57
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v58
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v59
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB34_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v59
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v39
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v58
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v3
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v57
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v62
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v47
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v46
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v45
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v61
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v43
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v60
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v41
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v63
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v11
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v40
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v55
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v53
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v52
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v51
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v50
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v49
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v27
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: .LBB34_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB34_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v32
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v58
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v56
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v47
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v62
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v46
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v45
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v37
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v44
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v43
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v36
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v42
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v41
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v61
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v40
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v55
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; SI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v35
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v54
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v53
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v60
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v52
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v51
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v34
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v49
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v48
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v63
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v33
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v29, 0x38000000, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v16
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v18, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v19
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v21, v19
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v27, v25
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_or_b32_e32 v24, v26, v24
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v27, v29, v27
|
|
; SI-NEXT: .LBB34_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56f16_to_v28f32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_mov_b32_e32 v32, v27
|
|
; VI-NEXT: v_mov_b32_e32 v33, v26
|
|
; VI-NEXT: v_mov_b32_e32 v34, v25
|
|
; VI-NEXT: v_mov_b32_e32 v35, v24
|
|
; VI-NEXT: v_mov_b32_e32 v36, v23
|
|
; VI-NEXT: v_mov_b32_e32 v37, v22
|
|
; VI-NEXT: v_mov_b32_e32 v38, v21
|
|
; VI-NEXT: v_mov_b32_e32 v39, v20
|
|
; VI-NEXT: v_mov_b32_e32 v48, v19
|
|
; VI-NEXT: v_mov_b32_e32 v49, v18
|
|
; VI-NEXT: v_mov_b32_e32 v50, v17
|
|
; VI-NEXT: v_mov_b32_e32 v51, v16
|
|
; VI-NEXT: v_mov_b32_e32 v52, v15
|
|
; VI-NEXT: v_mov_b32_e32 v53, v14
|
|
; VI-NEXT: v_mov_b32_e32 v54, v13
|
|
; VI-NEXT: v_mov_b32_e32 v55, v12
|
|
; VI-NEXT: v_mov_b32_e32 v40, v11
|
|
; VI-NEXT: v_mov_b32_e32 v41, v10
|
|
; VI-NEXT: v_mov_b32_e32 v42, v9
|
|
; VI-NEXT: v_mov_b32_e32 v43, v8
|
|
; VI-NEXT: v_mov_b32_e32 v44, v7
|
|
; VI-NEXT: v_mov_b32_e32 v45, v6
|
|
; VI-NEXT: v_mov_b32_e32 v46, v5
|
|
; VI-NEXT: v_mov_b32_e32 v47, v4
|
|
; VI-NEXT: v_mov_b32_e32 v56, v3
|
|
; VI-NEXT: v_mov_b32_e32 v57, v2
|
|
; VI-NEXT: v_mov_b32_e32 v58, v1
|
|
; VI-NEXT: v_mov_b32_e32 v59, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB34_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_mov_b32_e32 v27, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v27, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v27, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v27, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v27, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v27, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v27, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v27, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v27, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v27, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v27, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v27, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v27, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v27, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v27, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v27, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v27, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v27, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v27, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v27, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v27, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v27, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v24, v27, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v25, v27, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v26, v27, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v27, v27, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v56, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v47, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v46, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v45, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v44, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v43, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v42, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v41, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v40, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v53, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v52, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v51, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v50, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v49, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v39, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v37, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v36, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v35, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v34, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v33, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v32, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: ; implicit-def: $vgpr59
|
|
; VI-NEXT: ; implicit-def: $vgpr58
|
|
; VI-NEXT: ; implicit-def: $vgpr57
|
|
; VI-NEXT: ; implicit-def: $vgpr56
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: .LBB34_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB34_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; VI-NEXT: v_add_f16_sdwa v0, v59, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v59
|
|
; VI-NEXT: v_add_f16_sdwa v2, v58, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v58
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v2, v57, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v57
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v3, v56, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v56
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_add_f16_sdwa v4, v47, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v47
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_add_f16_sdwa v5, v46, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v46
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_add_f16_sdwa v6, v45, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v45
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_add_f16_sdwa v7, v44, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v44
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_add_f16_sdwa v8, v43, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v43
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_add_f16_sdwa v9, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, 0x200, v42
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v41, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, 0x200, v41
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_add_f16_sdwa v11, v40, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, 0x200, v40
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_add_f16_sdwa v12, v55, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, 0x200, v55
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_add_f16_sdwa v13, v54, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v54
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_add_f16_sdwa v14, v53, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v53
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v52, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v52
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v51, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v51
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v50, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v50
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v49, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v48, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v39, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v38, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v37, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v23, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v36, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: v_add_f16_sdwa v24, v35, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v25, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; VI-NEXT: v_add_f16_sdwa v25, v34, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v26, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; VI-NEXT: v_add_f16_sdwa v26, v33, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v26, v28, v26
|
|
; VI-NEXT: v_add_f16_sdwa v27, v32, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; VI-NEXT: .LBB34_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56f16_to_v28f32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_mov_b32_e32 v59, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v22
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v13
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v12
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v11
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v10
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v40
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v56, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v57, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v58, v1
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v41
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v47
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v56
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v57
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v58
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v59
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB34_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v37
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: .LBB34_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB34_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: s_movk_i32 s7, 0x200
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v0, v0, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, v1, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, v2, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, v3, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, v4, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, v5, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, v12, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, v13, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, v14, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, v15, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, v16, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, v17, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, v18, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, v19, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, v20, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, v21, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, v22, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, v23, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v24, v24, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v25, v25, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v26, v26, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v27, v27, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB34_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v28f32:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB34_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: .LBB34_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56f16_to_v28f32:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v69, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v71, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v80, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v68, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v67, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v66, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v65, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v64, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v54, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v53, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v52, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v51, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v50, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v49, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v48, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v39, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v38, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v37, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v36, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v35, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v34, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v33, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v32, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v31, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v30, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v29, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB34_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: .LBB34_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <56 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <56 x half> %a1 to <28 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x half> %a to <28 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x float> %phi
|
|
}
|
|
|
|
define inreg <28 x float> @bitcast_v56f16_to_v28f32_scalar(<56 x half> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v56f16_to_v28f32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v32, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v32, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v32, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v32, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v32, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v32, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v32, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v32, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v32, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v32, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v32, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v32, s51, 11
|
|
; SI-NEXT: v_writelane_b32 v32, s52, 12
|
|
; SI-NEXT: v_writelane_b32 v32, s53, 13
|
|
; SI-NEXT: v_writelane_b32 v32, s54, 14
|
|
; SI-NEXT: v_writelane_b32 v32, s55, 15
|
|
; SI-NEXT: v_writelane_b32 v32, s64, 16
|
|
; SI-NEXT: v_writelane_b32 v32, s65, 17
|
|
; SI-NEXT: v_writelane_b32 v32, s66, 18
|
|
; SI-NEXT: v_writelane_b32 v32, s67, 19
|
|
; SI-NEXT: v_writelane_b32 v32, s68, 20
|
|
; SI-NEXT: v_writelane_b32 v32, s69, 21
|
|
; SI-NEXT: v_writelane_b32 v32, s70, 22
|
|
; SI-NEXT: v_writelane_b32 v32, s71, 23
|
|
; SI-NEXT: v_writelane_b32 v32, s80, 24
|
|
; SI-NEXT: v_writelane_b32 v32, s81, 25
|
|
; SI-NEXT: v_writelane_b32 v32, s82, 26
|
|
; SI-NEXT: v_writelane_b32 v32, s83, 27
|
|
; SI-NEXT: v_writelane_b32 v32, s84, 28
|
|
; SI-NEXT: v_writelane_b32 v32, s85, 29
|
|
; SI-NEXT: v_writelane_b32 v32, s86, 30
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s72, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s74, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s76, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s78, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s88, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s90, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s92, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s95, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s34, v0
|
|
; SI-NEXT: v_writelane_b32 v32, s87, 31
|
|
; SI-NEXT: s_lshr_b32 s94, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s69, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s70, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s71, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s80, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s81, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s82, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s83, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s84, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s85, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s86, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s87, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s7, s6, 16
|
|
; SI-NEXT: s_lshr_b32 s9, s8, 16
|
|
; SI-NEXT: s_lshr_b32 s11, s10, 16
|
|
; SI-NEXT: s_lshr_b32 s13, s12, 16
|
|
; SI-NEXT: s_lshr_b32 s15, s14, 16
|
|
; SI-NEXT: s_lshr_b32 s73, s72, 16
|
|
; SI-NEXT: s_lshr_b32 s75, s74, 16
|
|
; SI-NEXT: s_lshr_b32 s77, s76, 16
|
|
; SI-NEXT: s_lshr_b32 s79, s78, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s88, 16
|
|
; SI-NEXT: s_lshr_b32 s91, s90, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s92, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s95, 16
|
|
; SI-NEXT: s_lshr_b32 s68, s34, 16
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB35_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s36, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s37, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s38, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s39, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s40, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s41, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s42, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; SI-NEXT: s_or_b32 s43, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s44, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s45, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s69, 16
|
|
; SI-NEXT: s_or_b32 s46, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s47, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s48, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s94, 16
|
|
; SI-NEXT: s_or_b32 s49, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s34, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s50, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s95, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; SI-NEXT: s_or_b32 s51, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s92, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s52, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s90, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s91, 16
|
|
; SI-NEXT: s_or_b32 s53, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s54, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s78, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s55, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s76, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s77, 16
|
|
; SI-NEXT: s_or_b32 s56, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s74, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; SI-NEXT: s_or_b32 s57, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s72, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s73, 16
|
|
; SI-NEXT: s_or_b32 s58, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; SI-NEXT: s_or_b32 s59, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; SI-NEXT: s_or_b32 s60, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; SI-NEXT: s_or_b32 s61, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s9, 16
|
|
; SI-NEXT: s_or_b32 s62, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; SI-NEXT: s_or_b32 s63, s4, s5
|
|
; SI-NEXT: s_cbranch_execnz .LBB35_4
|
|
; SI-NEXT: .LBB35_2: ; %cmp.true
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, s87
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s86
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s17
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s85
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s18
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s84
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s19
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s83
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s20
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_or_b32_e32 v3, v5, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s82
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_or_b32_e32 v4, v6, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s81
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s22
|
|
; SI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s80
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s23
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_or_b32_e32 v6, v8, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s71
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v9, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s70
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s25
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s69
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s26
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_or_b32_e32 v9, v11, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v12, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s30
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s28
|
|
; SI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s94
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s29
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s68
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v13, v15, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s31
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s95
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s93
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s92
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s91
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v18, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s90
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s89
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s88
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s79
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s78
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s77
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v21, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s76
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s75
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s74
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s73
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s72
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s13
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s12
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s10
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: v_or_b32_e32 v24, v26, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v27, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, s7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, s6
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v29, 0x38000000, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v27, v29, v27
|
|
; SI-NEXT: s_branch .LBB35_5
|
|
; SI-NEXT: .LBB35_3:
|
|
; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; SI-NEXT: s_branch .LBB35_2
|
|
; SI-NEXT: .LBB35_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_mov_b32_e32 v28, s64
|
|
; SI-NEXT: v_mov_b32_e32 v29, s65
|
|
; SI-NEXT: v_mov_b32_e32 v30, s66
|
|
; SI-NEXT: v_mov_b32_e32 v31, s67
|
|
; SI-NEXT: .LBB35_5: ; %end
|
|
; SI-NEXT: v_readlane_b32 s87, v32, 31
|
|
; SI-NEXT: v_readlane_b32 s86, v32, 30
|
|
; SI-NEXT: v_readlane_b32 s85, v32, 29
|
|
; SI-NEXT: v_readlane_b32 s84, v32, 28
|
|
; SI-NEXT: v_readlane_b32 s83, v32, 27
|
|
; SI-NEXT: v_readlane_b32 s82, v32, 26
|
|
; SI-NEXT: v_readlane_b32 s81, v32, 25
|
|
; SI-NEXT: v_readlane_b32 s80, v32, 24
|
|
; SI-NEXT: v_readlane_b32 s71, v32, 23
|
|
; SI-NEXT: v_readlane_b32 s70, v32, 22
|
|
; SI-NEXT: v_readlane_b32 s69, v32, 21
|
|
; SI-NEXT: v_readlane_b32 s68, v32, 20
|
|
; SI-NEXT: v_readlane_b32 s67, v32, 19
|
|
; SI-NEXT: v_readlane_b32 s66, v32, 18
|
|
; SI-NEXT: v_readlane_b32 s65, v32, 17
|
|
; SI-NEXT: v_readlane_b32 s64, v32, 16
|
|
; SI-NEXT: v_readlane_b32 s55, v32, 15
|
|
; SI-NEXT: v_readlane_b32 s54, v32, 14
|
|
; SI-NEXT: v_readlane_b32 s53, v32, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v32, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v32, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v32, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v32, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v32, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v32, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v32, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v32, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v32, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v32, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v32, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v32, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v32, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56f16_to_v28f32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v32, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s31, 1
|
|
; VI-NEXT: v_writelane_b32 v32, s34, 2
|
|
; VI-NEXT: v_writelane_b32 v32, s35, 3
|
|
; VI-NEXT: v_writelane_b32 v32, s36, 4
|
|
; VI-NEXT: v_writelane_b32 v32, s37, 5
|
|
; VI-NEXT: v_writelane_b32 v32, s38, 6
|
|
; VI-NEXT: v_writelane_b32 v32, s39, 7
|
|
; VI-NEXT: v_writelane_b32 v32, s48, 8
|
|
; VI-NEXT: v_writelane_b32 v32, s49, 9
|
|
; VI-NEXT: v_writelane_b32 v32, s50, 10
|
|
; VI-NEXT: v_writelane_b32 v32, s51, 11
|
|
; VI-NEXT: v_writelane_b32 v32, s52, 12
|
|
; VI-NEXT: v_writelane_b32 v32, s53, 13
|
|
; VI-NEXT: v_writelane_b32 v32, s54, 14
|
|
; VI-NEXT: v_writelane_b32 v32, s55, 15
|
|
; VI-NEXT: v_writelane_b32 v32, s64, 16
|
|
; VI-NEXT: v_writelane_b32 v32, s65, 17
|
|
; VI-NEXT: v_writelane_b32 v32, s66, 18
|
|
; VI-NEXT: v_writelane_b32 v32, s67, 19
|
|
; VI-NEXT: v_writelane_b32 v32, s68, 20
|
|
; VI-NEXT: v_writelane_b32 v32, s69, 21
|
|
; VI-NEXT: v_writelane_b32 v32, s70, 22
|
|
; VI-NEXT: v_writelane_b32 v32, s71, 23
|
|
; VI-NEXT: v_writelane_b32 v32, s80, 24
|
|
; VI-NEXT: v_writelane_b32 v32, s81, 25
|
|
; VI-NEXT: v_writelane_b32 v32, s82, 26
|
|
; VI-NEXT: v_writelane_b32 v32, s83, 27
|
|
; VI-NEXT: v_writelane_b32 v32, s84, 28
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; VI-NEXT: v_writelane_b32 v32, s85, 29
|
|
; VI-NEXT: s_lshr_b32 s15, s8, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v11
|
|
; VI-NEXT: ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
|
|
; VI-NEXT: v_writelane_b32 v32, s86, 30
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: s_lshr_b32 s61, s10, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s72, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s74, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s76, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s79, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s34, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s69, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s80, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s83, v0
|
|
; VI-NEXT: v_writelane_b32 v33, s15, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s87, 31
|
|
; VI-NEXT: s_lshr_b32 s56, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s68, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s70, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s81, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s84, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s86, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s87, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s7, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s9, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s11, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s13, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s72, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s74, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s76, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s79, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s91, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s34, 16
|
|
; VI-NEXT: s_lshr_b32 s71, s69, 16
|
|
; VI-NEXT: s_lshr_b32 s82, s80, 16
|
|
; VI-NEXT: s_lshr_b32 s85, s83, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v33, s61, 1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_writelane_b32 v33, s60, 2
|
|
; VI-NEXT: v_writelane_b32 v33, s59, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB35_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_or_b32 s36, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s5, s9, 16
|
|
; VI-NEXT: s_or_b32 s38, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s11, 16
|
|
; VI-NEXT: s_or_b32 s39, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; VI-NEXT: s_or_b32 s37, s40, s41
|
|
; VI-NEXT: s_or_b32 s40, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; VI-NEXT: s_or_b32 s41, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; VI-NEXT: s_or_b32 s42, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; VI-NEXT: s_or_b32 s43, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s44, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s45, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s46, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s47, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_or_b32 s48, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s5, s56, 16
|
|
; VI-NEXT: s_or_b32 s49, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s83
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s50, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s80
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s51, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s69
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s52, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s34
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s53, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s91
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s54, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s79
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s55, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s76
|
|
; VI-NEXT: s_lshl_b32 s5, s62, 16
|
|
; VI-NEXT: s_mov_b32 s77, s56
|
|
; VI-NEXT: s_or_b32 s56, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s74
|
|
; VI-NEXT: s_lshl_b32 s5, s57, 16
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: s_mov_b32 s75, s57
|
|
; VI-NEXT: s_or_b32 s57, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s72
|
|
; VI-NEXT: s_lshl_b32 s5, s58, 16
|
|
; VI-NEXT: s_mov_b32 s78, s58
|
|
; VI-NEXT: s_or_b32 s58, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s5, s59, 16
|
|
; VI-NEXT: s_or_b32 s59, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s5, s60, 16
|
|
; VI-NEXT: s_or_b32 s60, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s5, s61, 16
|
|
; VI-NEXT: s_or_b32 s61, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_mov_b32 s88, s62
|
|
; VI-NEXT: s_or_b32 s62, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s5, s63, 16
|
|
; VI-NEXT: s_mov_b32 s15, s63
|
|
; VI-NEXT: s_or_b32 s63, s4, s5
|
|
; VI-NEXT: s_cbranch_execnz .LBB35_4
|
|
; VI-NEXT: .LBB35_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; VI-NEXT: v_mov_b32_e32 v0, s13
|
|
; VI-NEXT: v_mov_b32_e32 v2, s11
|
|
; VI-NEXT: v_add_f16_sdwa v0, v0, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, s16, v27
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s17, v27
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v2, s9
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s18, v27
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v3, s7
|
|
; VI-NEXT: v_add_f16_sdwa v3, v3, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, s19, v27
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_mov_b32_e32 v4, s87
|
|
; VI-NEXT: v_add_f16_sdwa v4, v4, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, s20, v27
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_mov_b32_e32 v5, s86
|
|
; VI-NEXT: v_add_f16_sdwa v5, v5, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, s21, v27
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_mov_b32_e32 v6, s84
|
|
; VI-NEXT: v_add_f16_sdwa v6, v6, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, s22, v27
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_mov_b32_e32 v7, s81
|
|
; VI-NEXT: v_add_f16_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, s23, v27
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_mov_b32_e32 v8, s70
|
|
; VI-NEXT: v_add_f16_sdwa v8, v8, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, s24, v27
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_mov_b32_e32 v9, s68
|
|
; VI-NEXT: v_add_f16_sdwa v9, v9, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, s25, v27
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_mov_b32_e32 v10, s31
|
|
; VI-NEXT: v_add_f16_sdwa v10, v10, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, s26, v27
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_mov_b32_e32 v11, s90
|
|
; VI-NEXT: v_add_f16_sdwa v11, v11, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, s27, v27
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_mov_b32_e32 v12, s73
|
|
; VI-NEXT: v_add_f16_sdwa v12, v12, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, s28, v27
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_mov_b32_e32 v13, s77
|
|
; VI-NEXT: v_add_f16_sdwa v13, v13, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, s29, v27
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_mov_b32_e32 v14, s85
|
|
; VI-NEXT: v_add_f16_sdwa v14, v14, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, s83, v27
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_mov_b32_e32 v15, s82
|
|
; VI-NEXT: v_add_f16_sdwa v15, v15, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, s80, v27
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_mov_b32_e32 v16, s71
|
|
; VI-NEXT: v_add_f16_sdwa v16, v16, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, s69, v27
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_mov_b32_e32 v17, s35
|
|
; VI-NEXT: v_add_f16_sdwa v17, v17, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, s34, v27
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_mov_b32_e32 v18, s30
|
|
; VI-NEXT: v_add_f16_sdwa v18, v18, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, s91, v27
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_mov_b32_e32 v19, s89
|
|
; VI-NEXT: v_add_f16_sdwa v19, v19, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, s79, v27
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_mov_b32_e32 v20, s88
|
|
; VI-NEXT: v_add_f16_sdwa v20, v20, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, s76, v27
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_mov_b32_e32 v21, s75
|
|
; VI-NEXT: v_add_f16_sdwa v21, v21, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, s74, v27
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_mov_b32_e32 v22, s78
|
|
; VI-NEXT: v_add_f16_sdwa v22, v22, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v23, s72, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 3
|
|
; VI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; VI-NEXT: v_mov_b32_e32 v23, s4
|
|
; VI-NEXT: v_add_f16_sdwa v23, v23, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, s14, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 2
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: v_mov_b32_e32 v24, s4
|
|
; VI-NEXT: v_add_f16_sdwa v24, v24, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v25, s12, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 1
|
|
; VI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; VI-NEXT: v_mov_b32_e32 v25, s4
|
|
; VI-NEXT: v_add_f16_sdwa v25, v25, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v26, s10, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 0
|
|
; VI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; VI-NEXT: v_mov_b32_e32 v26, s4
|
|
; VI-NEXT: v_add_f16_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, s8, v27
|
|
; VI-NEXT: v_or_b32_e32 v26, v28, v26
|
|
; VI-NEXT: v_mov_b32_e32 v28, s15
|
|
; VI-NEXT: v_add_f16_sdwa v28, v28, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v27, s6, v27
|
|
; VI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; VI-NEXT: s_branch .LBB35_5
|
|
; VI-NEXT: .LBB35_3:
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: s_mov_b32 s88, s62
|
|
; VI-NEXT: s_mov_b32 s77, s56
|
|
; VI-NEXT: s_mov_b32 s75, s57
|
|
; VI-NEXT: s_mov_b32 s78, s58
|
|
; VI-NEXT: s_mov_b32 s15, s63
|
|
; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; VI-NEXT: s_branch .LBB35_2
|
|
; VI-NEXT: .LBB35_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_mov_b32_e32 v28, s64
|
|
; VI-NEXT: v_mov_b32_e32 v29, s65
|
|
; VI-NEXT: v_mov_b32_e32 v30, s66
|
|
; VI-NEXT: v_mov_b32_e32 v31, s67
|
|
; VI-NEXT: .LBB35_5: ; %end
|
|
; VI-NEXT: v_readlane_b32 s87, v32, 31
|
|
; VI-NEXT: v_readlane_b32 s86, v32, 30
|
|
; VI-NEXT: v_readlane_b32 s85, v32, 29
|
|
; VI-NEXT: v_readlane_b32 s84, v32, 28
|
|
; VI-NEXT: v_readlane_b32 s83, v32, 27
|
|
; VI-NEXT: v_readlane_b32 s82, v32, 26
|
|
; VI-NEXT: v_readlane_b32 s81, v32, 25
|
|
; VI-NEXT: v_readlane_b32 s80, v32, 24
|
|
; VI-NEXT: v_readlane_b32 s71, v32, 23
|
|
; VI-NEXT: v_readlane_b32 s70, v32, 22
|
|
; VI-NEXT: v_readlane_b32 s69, v32, 21
|
|
; VI-NEXT: v_readlane_b32 s68, v32, 20
|
|
; VI-NEXT: v_readlane_b32 s67, v32, 19
|
|
; VI-NEXT: v_readlane_b32 s66, v32, 18
|
|
; VI-NEXT: v_readlane_b32 s65, v32, 17
|
|
; VI-NEXT: v_readlane_b32 s64, v32, 16
|
|
; VI-NEXT: v_readlane_b32 s55, v32, 15
|
|
; VI-NEXT: v_readlane_b32 s54, v32, 14
|
|
; VI-NEXT: v_readlane_b32 s53, v32, 13
|
|
; VI-NEXT: v_readlane_b32 s52, v32, 12
|
|
; VI-NEXT: v_readlane_b32 s51, v32, 11
|
|
; VI-NEXT: v_readlane_b32 s50, v32, 10
|
|
; VI-NEXT: v_readlane_b32 s49, v32, 9
|
|
; VI-NEXT: v_readlane_b32 s48, v32, 8
|
|
; VI-NEXT: v_readlane_b32 s39, v32, 7
|
|
; VI-NEXT: v_readlane_b32 s38, v32, 6
|
|
; VI-NEXT: v_readlane_b32 s37, v32, 5
|
|
; VI-NEXT: v_readlane_b32 s36, v32, 4
|
|
; VI-NEXT: v_readlane_b32 s35, v32, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v32, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v32, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v32, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56f16_to_v28f32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: s_lshr_b32 s12, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s13, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s14, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s15, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s88, v5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v3
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v2
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: s_lshr_b32 s4, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s5, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s6, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s7, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s8, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s9, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s10, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s63, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s62, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s61, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s60, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s59, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s58, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s57, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s56, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s88, 16
|
|
; GFX9-NEXT: s_lshr_b32 s17, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s18, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s19, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s20, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s21, s12, 16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_cmp_lg_u32 s42, 0
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s50, s12, s21
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s51, s13, s20
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s52, s14, s19
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s53, s15, s18
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s54, s16, s17
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s55, s88, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s56, s56, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s57, s57, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s58, s58, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s59, s59, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s60, s60, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s61, s61, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s62, s62, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s63, s63, s72
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB35_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB35_4
|
|
; GFX9-NEXT: .LBB35_2: ; %cmp.true
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; GFX9-NEXT: v_pk_add_f16 v0, s36, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, s37, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, s38, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, s39, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, s40, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, s41, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, s42, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, s43, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, s44, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v9, s45, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, s46, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, s47, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, s48, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, s49, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, s50, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, s51, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, s52, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, s53, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, s54, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, s55, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, s56, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, s57, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, s58, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, s59, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v24, s60, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v25, s61, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v26, s62, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v27, s63, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_branch .LBB35_5
|
|
; GFX9-NEXT: .LBB35_3:
|
|
; GFX9-NEXT: s_branch .LBB35_2
|
|
; GFX9-NEXT: .LBB35_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB35_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v56f16_to_v28f32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s44, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s74, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s76, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s77, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s78, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s88, v10
|
|
; GFX11-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s0, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s44, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s56, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s58, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s59, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s60, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s62, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s74, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s76, 16
|
|
; GFX11-NEXT: s_lshr_b32 s90, s77, 16
|
|
; GFX11-NEXT: s_lshr_b32 s91, s78, 16
|
|
; GFX11-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s88, 0
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s78, s91
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s77, s90
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s76, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s74, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s62, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s60, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s59, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s58, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s56, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s44, s57
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB35_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB35_4
|
|
; GFX11-NEXT: .LBB35_2: ; %cmp.true
|
|
; GFX11-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v18, 0x200, s18 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v19, 0x200, s19 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v20, 0x200, s20 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v21, 0x200, s21 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v22, 0x200, s22 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v23, 0x200, s23 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v24, 0x200, s24 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v25, 0x200, s25 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v26, 0x200, s26 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v27, 0x200, s27 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB35_3:
|
|
; GFX11-NEXT: s_branch .LBB35_2
|
|
; GFX11-NEXT: .LBB35_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <56 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <56 x half> %a1 to <28 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x half> %a to <28 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <28 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <28 x float> %phi
|
|
}
|
|
|
|
define <14 x double> @bitcast_v14i64_to_v14f64(<14 x i64> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v14i64_to_v14f64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB36_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
|
|
; SI-NEXT: .LBB36_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14i64_to_v14f64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB36_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
|
|
; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
|
|
; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
|
|
; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
|
|
; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
|
|
; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
|
|
; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
|
|
; VI-NEXT: .LBB36_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14i64_to_v14f64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB36_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
|
|
; GFX9-NEXT: .LBB36_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v14i64_to_v14f64:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB36_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
|
|
; GFX11-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
|
|
; GFX11-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
|
|
; GFX11-NEXT: .LBB36_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <14 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <14 x i64> %a1 to <14 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x i64> %a to <14 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x double> %phi
|
|
}
|
|
|
|
define inreg <14 x double> @bitcast_v14i64_to_v14f64_scalar(<14 x i64> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v14i64_to_v14f64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB37_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB37_3
|
|
; SI-NEXT: .LBB37_2: ; %cmp.true
|
|
; SI-NEXT: s_add_u32 s16, s16, 3
|
|
; SI-NEXT: s_addc_u32 s17, s17, 0
|
|
; SI-NEXT: s_add_u32 s18, s18, 3
|
|
; SI-NEXT: s_addc_u32 s19, s19, 0
|
|
; SI-NEXT: s_add_u32 s20, s20, 3
|
|
; SI-NEXT: s_addc_u32 s21, s21, 0
|
|
; SI-NEXT: s_add_u32 s22, s22, 3
|
|
; SI-NEXT: s_addc_u32 s23, s23, 0
|
|
; SI-NEXT: s_add_u32 s24, s24, 3
|
|
; SI-NEXT: s_addc_u32 s25, s25, 0
|
|
; SI-NEXT: s_add_u32 s26, s26, 3
|
|
; SI-NEXT: s_addc_u32 s27, s27, 0
|
|
; SI-NEXT: s_add_u32 s28, s28, 3
|
|
; SI-NEXT: s_addc_u32 s29, s29, 0
|
|
; SI-NEXT: s_add_u32 s43, s43, 3
|
|
; SI-NEXT: s_addc_u32 s42, s42, 0
|
|
; SI-NEXT: s_add_u32 s41, s41, 3
|
|
; SI-NEXT: s_addc_u32 s40, s40, 0
|
|
; SI-NEXT: s_add_u32 s15, s15, 3
|
|
; SI-NEXT: s_addc_u32 s14, s14, 0
|
|
; SI-NEXT: s_add_u32 s13, s13, 3
|
|
; SI-NEXT: s_addc_u32 s12, s12, 0
|
|
; SI-NEXT: s_add_u32 s11, s11, 3
|
|
; SI-NEXT: s_addc_u32 s10, s10, 0
|
|
; SI-NEXT: s_add_u32 s9, s9, 3
|
|
; SI-NEXT: s_addc_u32 s8, s8, 0
|
|
; SI-NEXT: s_add_u32 s7, s7, 3
|
|
; SI-NEXT: s_addc_u32 s6, s6, 0
|
|
; SI-NEXT: .LBB37_3: ; %end
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s43
|
|
; SI-NEXT: v_mov_b32_e32 v15, s42
|
|
; SI-NEXT: v_mov_b32_e32 v16, s41
|
|
; SI-NEXT: v_mov_b32_e32 v17, s40
|
|
; SI-NEXT: v_mov_b32_e32 v18, s15
|
|
; SI-NEXT: v_mov_b32_e32 v19, s14
|
|
; SI-NEXT: v_mov_b32_e32 v20, s13
|
|
; SI-NEXT: v_mov_b32_e32 v21, s12
|
|
; SI-NEXT: v_mov_b32_e32 v22, s11
|
|
; SI-NEXT: v_mov_b32_e32 v23, s10
|
|
; SI-NEXT: v_mov_b32_e32 v24, s9
|
|
; SI-NEXT: v_mov_b32_e32 v25, s8
|
|
; SI-NEXT: v_mov_b32_e32 v26, s7
|
|
; SI-NEXT: v_mov_b32_e32 v27, s6
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB37_4:
|
|
; SI-NEXT: s_branch .LBB37_2
|
|
;
|
|
; VI-LABEL: bitcast_v14i64_to_v14f64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB37_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB37_3
|
|
; VI-NEXT: .LBB37_2: ; %cmp.true
|
|
; VI-NEXT: s_add_u32 s16, s16, 3
|
|
; VI-NEXT: s_addc_u32 s17, s17, 0
|
|
; VI-NEXT: s_add_u32 s18, s18, 3
|
|
; VI-NEXT: s_addc_u32 s19, s19, 0
|
|
; VI-NEXT: s_add_u32 s20, s20, 3
|
|
; VI-NEXT: s_addc_u32 s21, s21, 0
|
|
; VI-NEXT: s_add_u32 s22, s22, 3
|
|
; VI-NEXT: s_addc_u32 s23, s23, 0
|
|
; VI-NEXT: s_add_u32 s24, s24, 3
|
|
; VI-NEXT: s_addc_u32 s25, s25, 0
|
|
; VI-NEXT: s_add_u32 s26, s26, 3
|
|
; VI-NEXT: s_addc_u32 s27, s27, 0
|
|
; VI-NEXT: s_add_u32 s28, s28, 3
|
|
; VI-NEXT: s_addc_u32 s29, s29, 0
|
|
; VI-NEXT: s_add_u32 s43, s43, 3
|
|
; VI-NEXT: s_addc_u32 s42, s42, 0
|
|
; VI-NEXT: s_add_u32 s41, s41, 3
|
|
; VI-NEXT: s_addc_u32 s40, s40, 0
|
|
; VI-NEXT: s_add_u32 s15, s15, 3
|
|
; VI-NEXT: s_addc_u32 s14, s14, 0
|
|
; VI-NEXT: s_add_u32 s13, s13, 3
|
|
; VI-NEXT: s_addc_u32 s12, s12, 0
|
|
; VI-NEXT: s_add_u32 s11, s11, 3
|
|
; VI-NEXT: s_addc_u32 s10, s10, 0
|
|
; VI-NEXT: s_add_u32 s9, s9, 3
|
|
; VI-NEXT: s_addc_u32 s8, s8, 0
|
|
; VI-NEXT: s_add_u32 s7, s7, 3
|
|
; VI-NEXT: s_addc_u32 s6, s6, 0
|
|
; VI-NEXT: .LBB37_3: ; %end
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v14, s43
|
|
; VI-NEXT: v_mov_b32_e32 v15, s42
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB37_4:
|
|
; VI-NEXT: s_branch .LBB37_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v14i64_to_v14f64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB37_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB37_3
|
|
; GFX9-NEXT: .LBB37_2: ; %cmp.true
|
|
; GFX9-NEXT: s_add_u32 s16, s16, 3
|
|
; GFX9-NEXT: s_addc_u32 s17, s17, 0
|
|
; GFX9-NEXT: s_add_u32 s18, s18, 3
|
|
; GFX9-NEXT: s_addc_u32 s19, s19, 0
|
|
; GFX9-NEXT: s_add_u32 s20, s20, 3
|
|
; GFX9-NEXT: s_addc_u32 s21, s21, 0
|
|
; GFX9-NEXT: s_add_u32 s22, s22, 3
|
|
; GFX9-NEXT: s_addc_u32 s23, s23, 0
|
|
; GFX9-NEXT: s_add_u32 s24, s24, 3
|
|
; GFX9-NEXT: s_addc_u32 s25, s25, 0
|
|
; GFX9-NEXT: s_add_u32 s26, s26, 3
|
|
; GFX9-NEXT: s_addc_u32 s27, s27, 0
|
|
; GFX9-NEXT: s_add_u32 s28, s28, 3
|
|
; GFX9-NEXT: s_addc_u32 s29, s29, 0
|
|
; GFX9-NEXT: s_add_u32 s43, s43, 3
|
|
; GFX9-NEXT: s_addc_u32 s42, s42, 0
|
|
; GFX9-NEXT: s_add_u32 s41, s41, 3
|
|
; GFX9-NEXT: s_addc_u32 s40, s40, 0
|
|
; GFX9-NEXT: s_add_u32 s15, s15, 3
|
|
; GFX9-NEXT: s_addc_u32 s14, s14, 0
|
|
; GFX9-NEXT: s_add_u32 s13, s13, 3
|
|
; GFX9-NEXT: s_addc_u32 s12, s12, 0
|
|
; GFX9-NEXT: s_add_u32 s11, s11, 3
|
|
; GFX9-NEXT: s_addc_u32 s10, s10, 0
|
|
; GFX9-NEXT: s_add_u32 s9, s9, 3
|
|
; GFX9-NEXT: s_addc_u32 s8, s8, 0
|
|
; GFX9-NEXT: s_add_u32 s7, s7, 3
|
|
; GFX9-NEXT: s_addc_u32 s6, s6, 0
|
|
; GFX9-NEXT: .LBB37_3: ; %end
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB37_4:
|
|
; GFX9-NEXT: s_branch .LBB37_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v14i64_to_v14f64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB37_4
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB37_3
|
|
; GFX11-NEXT: .LBB37_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_u32 s0, s0, 3
|
|
; GFX11-NEXT: s_addc_u32 s1, s1, 0
|
|
; GFX11-NEXT: s_add_u32 s2, s2, 3
|
|
; GFX11-NEXT: s_addc_u32 s3, s3, 0
|
|
; GFX11-NEXT: s_add_u32 s16, s16, 3
|
|
; GFX11-NEXT: s_addc_u32 s17, s17, 0
|
|
; GFX11-NEXT: s_add_u32 s18, s18, 3
|
|
; GFX11-NEXT: s_addc_u32 s19, s19, 0
|
|
; GFX11-NEXT: s_add_u32 s20, s20, 3
|
|
; GFX11-NEXT: s_addc_u32 s21, s21, 0
|
|
; GFX11-NEXT: s_add_u32 s22, s22, 3
|
|
; GFX11-NEXT: s_addc_u32 s23, s23, 0
|
|
; GFX11-NEXT: s_add_u32 s24, s24, 3
|
|
; GFX11-NEXT: s_addc_u32 s25, s25, 0
|
|
; GFX11-NEXT: s_add_u32 s26, s26, 3
|
|
; GFX11-NEXT: s_addc_u32 s27, s27, 0
|
|
; GFX11-NEXT: s_add_u32 s28, s28, 3
|
|
; GFX11-NEXT: s_addc_u32 s29, s29, 0
|
|
; GFX11-NEXT: s_add_u32 s13, s13, 3
|
|
; GFX11-NEXT: s_addc_u32 s12, s12, 0
|
|
; GFX11-NEXT: s_add_u32 s11, s11, 3
|
|
; GFX11-NEXT: s_addc_u32 s10, s10, 0
|
|
; GFX11-NEXT: s_add_u32 s9, s9, 3
|
|
; GFX11-NEXT: s_addc_u32 s8, s8, 0
|
|
; GFX11-NEXT: s_add_u32 s7, s7, 3
|
|
; GFX11-NEXT: s_addc_u32 s6, s6, 0
|
|
; GFX11-NEXT: s_add_u32 s5, s5, 3
|
|
; GFX11-NEXT: s_addc_u32 s4, s4, 0
|
|
; GFX11-NEXT: .LBB37_3: ; %end
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB37_4:
|
|
; GFX11-NEXT: s_branch .LBB37_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <14 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <14 x i64> %a1 to <14 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x i64> %a to <14 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x double> %phi
|
|
}
|
|
|
|
define <14 x i64> @bitcast_v14f64_to_v14i64(<14 x double> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v14f64_to_v14i64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB38_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; SI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; SI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; SI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; SI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; SI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; SI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; SI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; SI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; SI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; SI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; SI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; SI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; SI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; SI-NEXT: .LBB38_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14f64_to_v14i64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB38_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; VI-NEXT: .LBB38_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14f64_to_v14i64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB38_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX9-NEXT: .LBB38_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v14f64_to_v14i64:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-NEXT: s_cbranch_execz .LBB38_2
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX11-NEXT: .LBB38_2: ; %end
|
|
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <14 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <14 x double> %a1 to <14 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x double> %a to <14 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x i64> %phi
|
|
}
|
|
|
|
define inreg <14 x i64> @bitcast_v14f64_to_v14i64_scalar(<14 x double> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v14f64_to_v14i64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; SI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; SI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; SI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; SI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; SI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; SI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; SI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; SI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; SI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; SI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; SI-NEXT: s_mov_b32 s49, s29
|
|
; SI-NEXT: s_mov_b32 s48, s28
|
|
; SI-NEXT: s_mov_b32 s47, s27
|
|
; SI-NEXT: s_mov_b32 s46, s26
|
|
; SI-NEXT: s_mov_b32 s45, s25
|
|
; SI-NEXT: s_mov_b32 s44, s24
|
|
; SI-NEXT: s_mov_b32 s43, s23
|
|
; SI-NEXT: s_mov_b32 s42, s22
|
|
; SI-NEXT: s_mov_b32 s41, s21
|
|
; SI-NEXT: s_mov_b32 s40, s20
|
|
; SI-NEXT: s_mov_b32 s39, s19
|
|
; SI-NEXT: s_mov_b32 s38, s18
|
|
; SI-NEXT: s_mov_b32 s37, s17
|
|
; SI-NEXT: s_mov_b32 s36, s16
|
|
; SI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB39_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB39_4
|
|
; SI-NEXT: .LBB39_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; SI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; SI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; SI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; SI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; SI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; SI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; SI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; SI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; SI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; SI-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; SI-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; SI-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; SI-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; SI-NEXT: s_branch .LBB39_5
|
|
; SI-NEXT: .LBB39_3:
|
|
; SI-NEXT: s_branch .LBB39_2
|
|
; SI-NEXT: .LBB39_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_mov_b32_e32 v28, s64
|
|
; SI-NEXT: v_mov_b32_e32 v29, s65
|
|
; SI-NEXT: v_mov_b32_e32 v30, s66
|
|
; SI-NEXT: v_mov_b32_e32 v31, s67
|
|
; SI-NEXT: .LBB39_5: ; %end
|
|
; SI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; SI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; SI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; SI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; SI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; SI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; SI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; SI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; SI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; SI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; SI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; SI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14f64_to_v14i64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v32, s36, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s37, 1
|
|
; VI-NEXT: v_writelane_b32 v32, s38, 2
|
|
; VI-NEXT: v_writelane_b32 v32, s39, 3
|
|
; VI-NEXT: v_writelane_b32 v32, s48, 4
|
|
; VI-NEXT: v_writelane_b32 v32, s49, 5
|
|
; VI-NEXT: v_writelane_b32 v32, s50, 6
|
|
; VI-NEXT: v_writelane_b32 v32, s51, 7
|
|
; VI-NEXT: v_writelane_b32 v32, s52, 8
|
|
; VI-NEXT: v_writelane_b32 v32, s53, 9
|
|
; VI-NEXT: v_writelane_b32 v32, s54, 10
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v32, s55, 11
|
|
; VI-NEXT: s_mov_b32 s49, s29
|
|
; VI-NEXT: s_mov_b32 s48, s28
|
|
; VI-NEXT: s_mov_b32 s47, s27
|
|
; VI-NEXT: s_mov_b32 s46, s26
|
|
; VI-NEXT: s_mov_b32 s45, s25
|
|
; VI-NEXT: s_mov_b32 s44, s24
|
|
; VI-NEXT: s_mov_b32 s43, s23
|
|
; VI-NEXT: s_mov_b32 s42, s22
|
|
; VI-NEXT: s_mov_b32 s41, s21
|
|
; VI-NEXT: s_mov_b32 s40, s20
|
|
; VI-NEXT: s_mov_b32 s39, s19
|
|
; VI-NEXT: s_mov_b32 s38, s18
|
|
; VI-NEXT: s_mov_b32 s37, s17
|
|
; VI-NEXT: s_mov_b32 s36, s16
|
|
; VI-NEXT: v_readfirstlane_b32 s63, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s62, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s61, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s60, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s59, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s58, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s57, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s56, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s55, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s54, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s53, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s52, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s51, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s50, v0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB39_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB39_4
|
|
; VI-NEXT: .LBB39_2: ; %cmp.true
|
|
; VI-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; VI-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; VI-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; VI-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; VI-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; VI-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; VI-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; VI-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; VI-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; VI-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; VI-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; VI-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; VI-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; VI-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; VI-NEXT: s_branch .LBB39_5
|
|
; VI-NEXT: .LBB39_3:
|
|
; VI-NEXT: s_branch .LBB39_2
|
|
; VI-NEXT: .LBB39_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_mov_b32_e32 v28, s64
|
|
; VI-NEXT: v_mov_b32_e32 v29, s65
|
|
; VI-NEXT: v_mov_b32_e32 v30, s66
|
|
; VI-NEXT: v_mov_b32_e32 v31, s67
|
|
; VI-NEXT: .LBB39_5: ; %end
|
|
; VI-NEXT: v_readlane_b32 s55, v32, 11
|
|
; VI-NEXT: v_readlane_b32 s54, v32, 10
|
|
; VI-NEXT: v_readlane_b32 s53, v32, 9
|
|
; VI-NEXT: v_readlane_b32 s52, v32, 8
|
|
; VI-NEXT: v_readlane_b32 s51, v32, 7
|
|
; VI-NEXT: v_readlane_b32 s50, v32, 6
|
|
; VI-NEXT: v_readlane_b32 s49, v32, 5
|
|
; VI-NEXT: v_readlane_b32 s48, v32, 4
|
|
; VI-NEXT: v_readlane_b32 s39, v32, 3
|
|
; VI-NEXT: v_readlane_b32 s38, v32, 2
|
|
; VI-NEXT: v_readlane_b32 s37, v32, 1
|
|
; VI-NEXT: v_readlane_b32 s36, v32, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14f64_to_v14i64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_mov_b32 s49, s29
|
|
; GFX9-NEXT: s_mov_b32 s48, s28
|
|
; GFX9-NEXT: s_mov_b32 s47, s27
|
|
; GFX9-NEXT: s_mov_b32 s46, s26
|
|
; GFX9-NEXT: s_mov_b32 s45, s25
|
|
; GFX9-NEXT: s_mov_b32 s44, s24
|
|
; GFX9-NEXT: s_mov_b32 s43, s23
|
|
; GFX9-NEXT: s_mov_b32 s42, s22
|
|
; GFX9-NEXT: s_mov_b32 s41, s21
|
|
; GFX9-NEXT: s_mov_b32 s40, s20
|
|
; GFX9-NEXT: s_mov_b32 s39, s19
|
|
; GFX9-NEXT: s_mov_b32 s38, s18
|
|
; GFX9-NEXT: s_mov_b32 s37, s17
|
|
; GFX9-NEXT: s_mov_b32 s36, s16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s55, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s54, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s53, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s52, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s51, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s50, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB39_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB39_4
|
|
; GFX9-NEXT: .LBB39_2: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; GFX9-NEXT: s_branch .LBB39_5
|
|
; GFX9-NEXT: .LBB39_3:
|
|
; GFX9-NEXT: s_branch .LBB39_2
|
|
; GFX9-NEXT: .LBB39_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB39_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v14f64_to_v14i64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s4, -1
|
|
; GFX11-NEXT: scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX11-NEXT: s_mov_b32 s36, s0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s0, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s63, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v8
|
|
; GFX11-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s61, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v4
|
|
; GFX11-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s57, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v2
|
|
; GFX11-NEXT: s_mov_b32 s47, s23
|
|
; GFX11-NEXT: s_mov_b32 s46, s22
|
|
; GFX11-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX11-NEXT: s_mov_b32 s45, s21
|
|
; GFX11-NEXT: s_mov_b32 s44, s20
|
|
; GFX11-NEXT: s_mov_b32 s43, s19
|
|
; GFX11-NEXT: s_mov_b32 s42, s18
|
|
; GFX11-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX11-NEXT: s_mov_b32 s48, s24
|
|
; GFX11-NEXT: s_mov_b32 s41, s17
|
|
; GFX11-NEXT: s_mov_b32 s40, s16
|
|
; GFX11-NEXT: s_mov_b32 s39, s3
|
|
; GFX11-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX11-NEXT: s_mov_b32 s49, s25
|
|
; GFX11-NEXT: s_mov_b32 s38, s2
|
|
; GFX11-NEXT: s_mov_b32 s37, s1
|
|
; GFX11-NEXT: s_cmp_lg_u32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX11-NEXT: s_mov_b32 s50, s26
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX11-NEXT: s_mov_b32 s51, s27
|
|
; GFX11-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX11-NEXT: s_mov_b32 s52, s28
|
|
; GFX11-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX11-NEXT: s_mov_b32 s53, s29
|
|
; GFX11-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s54, v0
|
|
; GFX11-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX11-NEXT: v_readfirstlane_b32 s55, v1
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB39_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB39_4
|
|
; GFX11-NEXT: .LBB39_2: ; %cmp.true
|
|
; GFX11-NEXT: v_add_f64 v[0:1], s[36:37], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[2:3], s[38:39], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[4:5], s[40:41], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[6:7], s[42:43], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[8:9], s[44:45], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[10:11], s[46:47], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[12:13], s[48:49], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[14:15], s[50:51], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[16:17], s[52:53], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[18:19], s[54:55], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[20:21], s[56:57], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[22:23], s[58:59], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[24:25], s[60:61], 1.0
|
|
; GFX11-NEXT: v_add_f64 v[26:27], s[62:63], 1.0
|
|
; GFX11-NEXT: s_branch .LBB39_5
|
|
; GFX11-NEXT: .LBB39_3:
|
|
; GFX11-NEXT: s_branch .LBB39_2
|
|
; GFX11-NEXT: .LBB39_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
|
|
; GFX11-NEXT: .LBB39_5: ; %end
|
|
; GFX11-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX11-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX11-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX11-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX11-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX11-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX11-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX11-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX11-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX11-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX11-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX11-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
|
|
; GFX11-NEXT: scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
|
|
; GFX11-NEXT: s_mov_b32 exec_lo, s0
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <14 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <14 x double> %a1 to <14 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x double> %a to <14 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x i64> %phi
|
|
}
|
|
|
|
define <56 x i16> @bitcast_v14i64_to_v56i16(<14 x i64> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v14i64_to_v56i16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB40_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v37, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v39, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB40_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB40_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v37, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v39, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB40_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v47
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v40
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v41
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v53
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v50
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v39
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v37
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14i64_to_v56i16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB40_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB40_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB40_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
|
|
; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
|
|
; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
|
|
; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
|
|
; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
|
|
; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB40_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14i64_to_v56i16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB40_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB40_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB40_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB40_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v47, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v46, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v45, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v44, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v43, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v42, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v41, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v40, v7, s4
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v8, v55, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v54, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v53, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v52, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v51, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v49, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v48, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v39, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v38, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v37, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v36, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v35, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v34, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v33, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v32, v23, s4
|
|
; GFX9-NEXT: v_perm_b32 v24, v31, v24, s4
|
|
; GFX9-NEXT: v_perm_b32 v25, v30, v25, s4
|
|
; GFX9-NEXT: v_perm_b32 v26, v29, v26, s4
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s4
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v14i64_to_v56i16:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB40_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
|
|
; GFX11-TRUE16-NEXT: .LBB40_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v14i64_to_v56i16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr67
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB40_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB40_2: ; %Flow
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB40_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB40_4: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v71, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v69, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v68, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v67, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v66, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v65, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v64, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v55, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v54, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v53, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v52, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v51, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v49, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v39, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v37, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v36, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v35, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v34, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v33, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v32, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v31, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v30, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v29, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <14 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <14 x i64> %a1 to <56 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x i64> %a to <56 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x i16> %phi
|
|
}
|
|
|
|
define inreg <56 x i16> @bitcast_v14i64_to_v56i16_scalar(<14 x i64> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v14i64_to_v56i16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v14
|
|
; SI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; SI-NEXT: v_readfirstlane_b32 s5, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s40, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v0
|
|
; SI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; SI-NEXT: s_cbranch_scc0 .LBB41_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s30, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s53, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB41_3
|
|
; SI-NEXT: .LBB41_2: ; %cmp.true
|
|
; SI-NEXT: s_add_u32 s4, s4, 3
|
|
; SI-NEXT: s_addc_u32 s5, s5, 0
|
|
; SI-NEXT: s_add_u32 s6, s6, 3
|
|
; SI-NEXT: s_addc_u32 s7, s7, 0
|
|
; SI-NEXT: s_add_u32 s8, s8, 3
|
|
; SI-NEXT: s_addc_u32 s9, s9, 0
|
|
; SI-NEXT: s_add_u32 s10, s10, 3
|
|
; SI-NEXT: s_addc_u32 s11, s11, 0
|
|
; SI-NEXT: s_add_u32 s12, s12, 3
|
|
; SI-NEXT: s_addc_u32 s13, s13, 0
|
|
; SI-NEXT: s_add_u32 s14, s14, 3
|
|
; SI-NEXT: s_addc_u32 s15, s15, 0
|
|
; SI-NEXT: s_add_u32 s40, s40, 3
|
|
; SI-NEXT: s_addc_u32 s41, s41, 0
|
|
; SI-NEXT: s_add_u32 s28, s28, 3
|
|
; SI-NEXT: s_addc_u32 s29, s29, 0
|
|
; SI-NEXT: s_add_u32 s26, s26, 3
|
|
; SI-NEXT: s_addc_u32 s27, s27, 0
|
|
; SI-NEXT: s_add_u32 s24, s24, 3
|
|
; SI-NEXT: s_addc_u32 s25, s25, 0
|
|
; SI-NEXT: s_add_u32 s22, s22, 3
|
|
; SI-NEXT: s_addc_u32 s23, s23, 0
|
|
; SI-NEXT: s_add_u32 s20, s20, 3
|
|
; SI-NEXT: s_addc_u32 s21, s21, 0
|
|
; SI-NEXT: s_add_u32 s18, s18, 3
|
|
; SI-NEXT: s_addc_u32 s19, s19, 0
|
|
; SI-NEXT: s_add_u32 s16, s16, 3
|
|
; SI-NEXT: s_addc_u32 s17, s17, 0
|
|
; SI-NEXT: s_lshr_b32 s30, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s53, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: .LBB41_3: ; %end
|
|
; SI-NEXT: s_lshl_b32 s43, s92, 16
|
|
; SI-NEXT: s_and_b32 s16, s16, 0xffff
|
|
; SI-NEXT: s_or_b32 s16, s16, s43
|
|
; SI-NEXT: s_and_b32 s17, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s53, 16
|
|
; SI-NEXT: s_or_b32 s17, s17, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s90, 16
|
|
; SI-NEXT: s_and_b32 s18, s18, 0xffff
|
|
; SI-NEXT: s_or_b32 s18, s18, s43
|
|
; SI-NEXT: s_and_b32 s19, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s52, 16
|
|
; SI-NEXT: s_or_b32 s19, s19, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s88, 16
|
|
; SI-NEXT: s_and_b32 s20, s20, 0xffff
|
|
; SI-NEXT: s_or_b32 s20, s20, s43
|
|
; SI-NEXT: s_and_b32 s21, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s51, 16
|
|
; SI-NEXT: s_or_b32 s21, s21, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s78, 16
|
|
; SI-NEXT: s_and_b32 s22, s22, 0xffff
|
|
; SI-NEXT: s_or_b32 s22, s22, s43
|
|
; SI-NEXT: s_and_b32 s23, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s50, 16
|
|
; SI-NEXT: s_or_b32 s23, s23, s43
|
|
; SI-NEXT: s_and_b32 s24, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s76, 16
|
|
; SI-NEXT: s_or_b32 s24, s24, s43
|
|
; SI-NEXT: s_and_b32 s25, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s49, 16
|
|
; SI-NEXT: s_or_b32 s25, s25, s43
|
|
; SI-NEXT: s_and_b32 s26, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s74, 16
|
|
; SI-NEXT: s_or_b32 s26, s26, s43
|
|
; SI-NEXT: s_and_b32 s27, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s48, 16
|
|
; SI-NEXT: s_or_b32 s27, s27, s43
|
|
; SI-NEXT: s_and_b32 s28, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s72, 16
|
|
; SI-NEXT: s_or_b32 s28, s28, s43
|
|
; SI-NEXT: s_and_b32 s29, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s39, 16
|
|
; SI-NEXT: s_or_b32 s29, s29, s43
|
|
; SI-NEXT: s_and_b32 s40, s40, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s62, 16
|
|
; SI-NEXT: s_or_b32 s40, s40, s43
|
|
; SI-NEXT: s_and_b32 s41, s41, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s38, 16
|
|
; SI-NEXT: s_or_b32 s41, s41, s43
|
|
; SI-NEXT: s_and_b32 s14, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s60, 16
|
|
; SI-NEXT: s_or_b32 s14, s14, s43
|
|
; SI-NEXT: s_and_b32 s15, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s37, 16
|
|
; SI-NEXT: s_or_b32 s15, s15, s43
|
|
; SI-NEXT: s_and_b32 s12, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s58, 16
|
|
; SI-NEXT: s_or_b32 s12, s12, s43
|
|
; SI-NEXT: s_and_b32 s13, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s36, 16
|
|
; SI-NEXT: s_or_b32 s13, s13, s43
|
|
; SI-NEXT: s_and_b32 s10, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s56, 16
|
|
; SI-NEXT: s_or_b32 s10, s10, s43
|
|
; SI-NEXT: s_and_b32 s11, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s35, 16
|
|
; SI-NEXT: s_or_b32 s11, s11, s43
|
|
; SI-NEXT: s_and_b32 s8, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s46, 16
|
|
; SI-NEXT: s_or_b32 s8, s8, s43
|
|
; SI-NEXT: s_and_b32 s9, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s34, 16
|
|
; SI-NEXT: s_or_b32 s9, s9, s43
|
|
; SI-NEXT: s_and_b32 s6, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s44, 16
|
|
; SI-NEXT: s_and_b32 s4, s4, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s42, 16
|
|
; SI-NEXT: s_or_b32 s6, s6, s43
|
|
; SI-NEXT: s_and_b32 s7, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s31, 16
|
|
; SI-NEXT: s_or_b32 s4, s4, s42
|
|
; SI-NEXT: s_and_b32 s5, s5, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s30, 16
|
|
; SI-NEXT: s_or_b32 s7, s7, s43
|
|
; SI-NEXT: s_or_b32 s5, s5, s42
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s40
|
|
; SI-NEXT: v_mov_b32_e32 v15, s41
|
|
; SI-NEXT: v_mov_b32_e32 v16, s14
|
|
; SI-NEXT: v_mov_b32_e32 v17, s15
|
|
; SI-NEXT: v_mov_b32_e32 v18, s12
|
|
; SI-NEXT: v_mov_b32_e32 v19, s13
|
|
; SI-NEXT: v_mov_b32_e32 v20, s10
|
|
; SI-NEXT: v_mov_b32_e32 v21, s11
|
|
; SI-NEXT: v_mov_b32_e32 v22, s8
|
|
; SI-NEXT: v_mov_b32_e32 v23, s9
|
|
; SI-NEXT: v_mov_b32_e32 v24, s6
|
|
; SI-NEXT: v_mov_b32_e32 v25, s7
|
|
; SI-NEXT: v_mov_b32_e32 v26, s4
|
|
; SI-NEXT: v_mov_b32_e32 v27, s5
|
|
; SI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB41_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr92
|
|
; SI-NEXT: ; implicit-def: $sgpr53
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: ; implicit-def: $sgpr52
|
|
; SI-NEXT: ; implicit-def: $sgpr88
|
|
; SI-NEXT: ; implicit-def: $sgpr51
|
|
; SI-NEXT: ; implicit-def: $sgpr78
|
|
; SI-NEXT: ; implicit-def: $sgpr50
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr49
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr48
|
|
; SI-NEXT: ; implicit-def: $sgpr72
|
|
; SI-NEXT: ; implicit-def: $sgpr39
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr38
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr37
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr36
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr35
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr31
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: s_branch .LBB41_2
|
|
;
|
|
; VI-LABEL: bitcast_v14i64_to_v56i16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB41_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s44, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s42, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB41_3
|
|
; VI-NEXT: .LBB41_2: ; %cmp.true
|
|
; VI-NEXT: s_add_u32 s7, s7, 3
|
|
; VI-NEXT: s_addc_u32 s6, s6, 0
|
|
; VI-NEXT: s_add_u32 s9, s9, 3
|
|
; VI-NEXT: s_addc_u32 s8, s8, 0
|
|
; VI-NEXT: s_add_u32 s11, s11, 3
|
|
; VI-NEXT: s_addc_u32 s10, s10, 0
|
|
; VI-NEXT: s_add_u32 s13, s13, 3
|
|
; VI-NEXT: s_addc_u32 s12, s12, 0
|
|
; VI-NEXT: s_add_u32 s15, s15, 3
|
|
; VI-NEXT: s_addc_u32 s14, s14, 0
|
|
; VI-NEXT: s_add_u32 s41, s41, 3
|
|
; VI-NEXT: s_addc_u32 s40, s40, 0
|
|
; VI-NEXT: s_add_u32 s43, s43, 3
|
|
; VI-NEXT: s_addc_u32 s42, s42, 0
|
|
; VI-NEXT: s_add_u32 s28, s28, 3
|
|
; VI-NEXT: s_addc_u32 s29, s29, 0
|
|
; VI-NEXT: s_add_u32 s26, s26, 3
|
|
; VI-NEXT: s_addc_u32 s27, s27, 0
|
|
; VI-NEXT: s_add_u32 s24, s24, 3
|
|
; VI-NEXT: s_addc_u32 s25, s25, 0
|
|
; VI-NEXT: s_add_u32 s22, s22, 3
|
|
; VI-NEXT: s_addc_u32 s23, s23, 0
|
|
; VI-NEXT: s_add_u32 s20, s20, 3
|
|
; VI-NEXT: s_addc_u32 s21, s21, 0
|
|
; VI-NEXT: s_add_u32 s18, s18, 3
|
|
; VI-NEXT: s_addc_u32 s19, s19, 0
|
|
; VI-NEXT: s_add_u32 s16, s16, 3
|
|
; VI-NEXT: s_addc_u32 s17, s17, 0
|
|
; VI-NEXT: s_lshr_b32 s44, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s42, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: .LBB41_3: ; %end
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s16, s34, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s16
|
|
; VI-NEXT: s_and_b32 s16, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s17, s31, 16
|
|
; VI-NEXT: s_or_b32 s16, s16, s17
|
|
; VI-NEXT: s_and_b32 s17, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s18, s30, 16
|
|
; VI-NEXT: s_or_b32 s17, s17, s18
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s19, s91, 16
|
|
; VI-NEXT: s_or_b32 s18, s18, s19
|
|
; VI-NEXT: s_and_b32 s19, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s20, s90, 16
|
|
; VI-NEXT: s_or_b32 s19, s19, s20
|
|
; VI-NEXT: s_and_b32 s20, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s21, s89, 16
|
|
; VI-NEXT: s_or_b32 s20, s20, s21
|
|
; VI-NEXT: s_and_b32 s21, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s22, s88, 16
|
|
; VI-NEXT: s_or_b32 s21, s21, s22
|
|
; VI-NEXT: s_and_b32 s22, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s23, s79, 16
|
|
; VI-NEXT: s_or_b32 s22, s22, s23
|
|
; VI-NEXT: s_and_b32 s23, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s24, s78, 16
|
|
; VI-NEXT: s_or_b32 s23, s23, s24
|
|
; VI-NEXT: s_and_b32 s24, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s25, s77, 16
|
|
; VI-NEXT: s_or_b32 s24, s24, s25
|
|
; VI-NEXT: s_and_b32 s25, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s26, s76, 16
|
|
; VI-NEXT: s_or_b32 s25, s25, s26
|
|
; VI-NEXT: s_and_b32 s26, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s27, s75, 16
|
|
; VI-NEXT: s_or_b32 s26, s26, s27
|
|
; VI-NEXT: s_and_b32 s27, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s28, s74, 16
|
|
; VI-NEXT: s_or_b32 s27, s27, s28
|
|
; VI-NEXT: s_and_b32 s28, 0xffff, s43
|
|
; VI-NEXT: s_lshl_b32 s29, s73, 16
|
|
; VI-NEXT: s_or_b32 s28, s28, s29
|
|
; VI-NEXT: s_and_b32 s29, 0xffff, s42
|
|
; VI-NEXT: s_lshl_b32 s42, s72, 16
|
|
; VI-NEXT: s_or_b32 s29, s29, s42
|
|
; VI-NEXT: s_and_b32 s41, 0xffff, s41
|
|
; VI-NEXT: s_lshl_b32 s42, s63, 16
|
|
; VI-NEXT: s_or_b32 s41, s41, s42
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s40
|
|
; VI-NEXT: s_lshl_b32 s42, s62, 16
|
|
; VI-NEXT: s_or_b32 s40, s40, s42
|
|
; VI-NEXT: s_and_b32 s15, 0xffff, s15
|
|
; VI-NEXT: s_lshl_b32 s42, s61, 16
|
|
; VI-NEXT: s_or_b32 s15, s15, s42
|
|
; VI-NEXT: s_and_b32 s14, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s42, s60, 16
|
|
; VI-NEXT: s_or_b32 s14, s14, s42
|
|
; VI-NEXT: s_and_b32 s13, 0xffff, s13
|
|
; VI-NEXT: s_lshl_b32 s42, s59, 16
|
|
; VI-NEXT: s_or_b32 s13, s13, s42
|
|
; VI-NEXT: s_and_b32 s12, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s42, s58, 16
|
|
; VI-NEXT: s_or_b32 s12, s12, s42
|
|
; VI-NEXT: s_and_b32 s11, 0xffff, s11
|
|
; VI-NEXT: s_lshl_b32 s42, s57, 16
|
|
; VI-NEXT: s_or_b32 s11, s11, s42
|
|
; VI-NEXT: s_and_b32 s10, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s42, s56, 16
|
|
; VI-NEXT: s_or_b32 s10, s10, s42
|
|
; VI-NEXT: s_and_b32 s9, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s42, s47, 16
|
|
; VI-NEXT: s_or_b32 s9, s9, s42
|
|
; VI-NEXT: s_and_b32 s8, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s42, s46, 16
|
|
; VI-NEXT: s_or_b32 s8, s8, s42
|
|
; VI-NEXT: s_and_b32 s7, 0xffff, s7
|
|
; VI-NEXT: s_lshl_b32 s42, s45, 16
|
|
; VI-NEXT: s_or_b32 s7, s7, s42
|
|
; VI-NEXT: s_and_b32 s6, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s42, s44, 16
|
|
; VI-NEXT: s_or_b32 s6, s6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s16
|
|
; VI-NEXT: v_mov_b32_e32 v3, s17
|
|
; VI-NEXT: v_mov_b32_e32 v4, s18
|
|
; VI-NEXT: v_mov_b32_e32 v5, s19
|
|
; VI-NEXT: v_mov_b32_e32 v6, s20
|
|
; VI-NEXT: v_mov_b32_e32 v7, s21
|
|
; VI-NEXT: v_mov_b32_e32 v8, s22
|
|
; VI-NEXT: v_mov_b32_e32 v9, s23
|
|
; VI-NEXT: v_mov_b32_e32 v10, s24
|
|
; VI-NEXT: v_mov_b32_e32 v11, s25
|
|
; VI-NEXT: v_mov_b32_e32 v12, s26
|
|
; VI-NEXT: v_mov_b32_e32 v13, s27
|
|
; VI-NEXT: v_mov_b32_e32 v14, s28
|
|
; VI-NEXT: v_mov_b32_e32 v15, s29
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB41_4:
|
|
; VI-NEXT: ; implicit-def: $sgpr35
|
|
; VI-NEXT: ; implicit-def: $sgpr34
|
|
; VI-NEXT: ; implicit-def: $sgpr31
|
|
; VI-NEXT: ; implicit-def: $sgpr30
|
|
; VI-NEXT: ; implicit-def: $sgpr91
|
|
; VI-NEXT: ; implicit-def: $sgpr90
|
|
; VI-NEXT: ; implicit-def: $sgpr89
|
|
; VI-NEXT: ; implicit-def: $sgpr88
|
|
; VI-NEXT: ; implicit-def: $sgpr79
|
|
; VI-NEXT: ; implicit-def: $sgpr78
|
|
; VI-NEXT: ; implicit-def: $sgpr77
|
|
; VI-NEXT: ; implicit-def: $sgpr76
|
|
; VI-NEXT: ; implicit-def: $sgpr75
|
|
; VI-NEXT: ; implicit-def: $sgpr74
|
|
; VI-NEXT: ; implicit-def: $sgpr73
|
|
; VI-NEXT: ; implicit-def: $sgpr72
|
|
; VI-NEXT: ; implicit-def: $sgpr63
|
|
; VI-NEXT: ; implicit-def: $sgpr62
|
|
; VI-NEXT: ; implicit-def: $sgpr61
|
|
; VI-NEXT: ; implicit-def: $sgpr60
|
|
; VI-NEXT: ; implicit-def: $sgpr59
|
|
; VI-NEXT: ; implicit-def: $sgpr58
|
|
; VI-NEXT: ; implicit-def: $sgpr57
|
|
; VI-NEXT: ; implicit-def: $sgpr56
|
|
; VI-NEXT: ; implicit-def: $sgpr47
|
|
; VI-NEXT: ; implicit-def: $sgpr46
|
|
; VI-NEXT: ; implicit-def: $sgpr45
|
|
; VI-NEXT: ; implicit-def: $sgpr44
|
|
; VI-NEXT: s_branch .LBB41_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v14i64_to_v56i16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB41_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s44, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s42, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s43, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB41_3
|
|
; GFX9-NEXT: .LBB41_2: ; %cmp.true
|
|
; GFX9-NEXT: s_add_u32 s7, s7, 3
|
|
; GFX9-NEXT: s_addc_u32 s6, s6, 0
|
|
; GFX9-NEXT: s_add_u32 s9, s9, 3
|
|
; GFX9-NEXT: s_addc_u32 s8, s8, 0
|
|
; GFX9-NEXT: s_add_u32 s11, s11, 3
|
|
; GFX9-NEXT: s_addc_u32 s10, s10, 0
|
|
; GFX9-NEXT: s_add_u32 s13, s13, 3
|
|
; GFX9-NEXT: s_addc_u32 s12, s12, 0
|
|
; GFX9-NEXT: s_add_u32 s15, s15, 3
|
|
; GFX9-NEXT: s_addc_u32 s14, s14, 0
|
|
; GFX9-NEXT: s_add_u32 s41, s41, 3
|
|
; GFX9-NEXT: s_addc_u32 s40, s40, 0
|
|
; GFX9-NEXT: s_add_u32 s43, s43, 3
|
|
; GFX9-NEXT: s_addc_u32 s42, s42, 0
|
|
; GFX9-NEXT: s_add_u32 s28, s28, 3
|
|
; GFX9-NEXT: s_addc_u32 s29, s29, 0
|
|
; GFX9-NEXT: s_add_u32 s26, s26, 3
|
|
; GFX9-NEXT: s_addc_u32 s27, s27, 0
|
|
; GFX9-NEXT: s_add_u32 s24, s24, 3
|
|
; GFX9-NEXT: s_addc_u32 s25, s25, 0
|
|
; GFX9-NEXT: s_add_u32 s22, s22, 3
|
|
; GFX9-NEXT: s_addc_u32 s23, s23, 0
|
|
; GFX9-NEXT: s_add_u32 s20, s20, 3
|
|
; GFX9-NEXT: s_addc_u32 s21, s21, 0
|
|
; GFX9-NEXT: s_add_u32 s18, s18, 3
|
|
; GFX9-NEXT: s_addc_u32 s19, s19, 0
|
|
; GFX9-NEXT: s_add_u32 s16, s16, 3
|
|
; GFX9-NEXT: s_addc_u32 s17, s17, 0
|
|
; GFX9-NEXT: s_lshr_b32 s44, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s42, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s43, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: .LBB41_3: ; %end
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s95
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s94
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s93
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s92
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s91
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s90
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s88
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s28, s43, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s29, s42, s72
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s41, s63
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s40, s62
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s15, s61
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s14, s60
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s13, s59
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s12, s58
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s57
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s56
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s47
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s46
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s45
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s4
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s5
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB41_4:
|
|
; GFX9-NEXT: ; implicit-def: $sgpr95
|
|
; GFX9-NEXT: ; implicit-def: $sgpr94
|
|
; GFX9-NEXT: ; implicit-def: $sgpr93
|
|
; GFX9-NEXT: ; implicit-def: $sgpr92
|
|
; GFX9-NEXT: ; implicit-def: $sgpr91
|
|
; GFX9-NEXT: ; implicit-def: $sgpr90
|
|
; GFX9-NEXT: ; implicit-def: $sgpr89
|
|
; GFX9-NEXT: ; implicit-def: $sgpr88
|
|
; GFX9-NEXT: ; implicit-def: $sgpr79
|
|
; GFX9-NEXT: ; implicit-def: $sgpr78
|
|
; GFX9-NEXT: ; implicit-def: $sgpr77
|
|
; GFX9-NEXT: ; implicit-def: $sgpr76
|
|
; GFX9-NEXT: ; implicit-def: $sgpr75
|
|
; GFX9-NEXT: ; implicit-def: $sgpr74
|
|
; GFX9-NEXT: ; implicit-def: $sgpr73
|
|
; GFX9-NEXT: ; implicit-def: $sgpr72
|
|
; GFX9-NEXT: ; implicit-def: $sgpr63
|
|
; GFX9-NEXT: ; implicit-def: $sgpr62
|
|
; GFX9-NEXT: ; implicit-def: $sgpr61
|
|
; GFX9-NEXT: ; implicit-def: $sgpr60
|
|
; GFX9-NEXT: ; implicit-def: $sgpr59
|
|
; GFX9-NEXT: ; implicit-def: $sgpr58
|
|
; GFX9-NEXT: ; implicit-def: $sgpr57
|
|
; GFX9-NEXT: ; implicit-def: $sgpr56
|
|
; GFX9-NEXT: ; implicit-def: $sgpr47
|
|
; GFX9-NEXT: ; implicit-def: $sgpr46
|
|
; GFX9-NEXT: ; implicit-def: $sgpr45
|
|
; GFX9-NEXT: ; implicit-def: $sgpr44
|
|
; GFX9-NEXT: s_branch .LBB41_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v14i64_to_v56i16_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-NEXT: s_mov_b32 s90, 0
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB41_4
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-NEXT: s_lshr_b32 s14, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s74, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s76, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s77, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s78, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s88, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s0, 16
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s90
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB41_3
|
|
; GFX11-NEXT: .LBB41_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_u32 s5, s5, 3
|
|
; GFX11-NEXT: s_addc_u32 s4, s4, 0
|
|
; GFX11-NEXT: s_add_u32 s7, s7, 3
|
|
; GFX11-NEXT: s_addc_u32 s6, s6, 0
|
|
; GFX11-NEXT: s_add_u32 s9, s9, 3
|
|
; GFX11-NEXT: s_addc_u32 s8, s8, 0
|
|
; GFX11-NEXT: s_add_u32 s11, s11, 3
|
|
; GFX11-NEXT: s_addc_u32 s10, s10, 0
|
|
; GFX11-NEXT: s_add_u32 s13, s13, 3
|
|
; GFX11-NEXT: s_addc_u32 s12, s12, 0
|
|
; GFX11-NEXT: s_add_u32 s28, s28, 3
|
|
; GFX11-NEXT: s_addc_u32 s29, s29, 0
|
|
; GFX11-NEXT: s_add_u32 s26, s26, 3
|
|
; GFX11-NEXT: s_addc_u32 s27, s27, 0
|
|
; GFX11-NEXT: s_add_u32 s24, s24, 3
|
|
; GFX11-NEXT: s_addc_u32 s25, s25, 0
|
|
; GFX11-NEXT: s_add_u32 s22, s22, 3
|
|
; GFX11-NEXT: s_addc_u32 s23, s23, 0
|
|
; GFX11-NEXT: s_add_u32 s20, s20, 3
|
|
; GFX11-NEXT: s_addc_u32 s21, s21, 0
|
|
; GFX11-NEXT: s_add_u32 s18, s18, 3
|
|
; GFX11-NEXT: s_addc_u32 s19, s19, 0
|
|
; GFX11-NEXT: s_add_u32 s16, s16, 3
|
|
; GFX11-NEXT: s_addc_u32 s17, s17, 0
|
|
; GFX11-NEXT: s_add_u32 s2, s2, 3
|
|
; GFX11-NEXT: s_addc_u32 s3, s3, 0
|
|
; GFX11-NEXT: s_add_u32 s0, s0, 3
|
|
; GFX11-NEXT: s_addc_u32 s1, s1, 0
|
|
; GFX11-NEXT: s_lshr_b32 s14, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s74, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s76, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s77, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s78, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s88, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s0, 16
|
|
; GFX11-NEXT: .LBB41_3: ; %end
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s88
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s78
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s77
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s76
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s74
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s62
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s60
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s59
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s58
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s57
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s56
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s44
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s40
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s14
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB41_4:
|
|
; GFX11-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-NEXT: ; implicit-def: $sgpr14
|
|
; GFX11-NEXT: s_branch .LBB41_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <14 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <14 x i64> %a1 to <56 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x i64> %a to <56 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x i16> %phi
|
|
}
|
|
|
|
define <14 x i64> @bitcast_v56i16_to_v14i64(<56 x i16> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v56i16_to_v14i64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_mov_b32_e32 v48, v19
|
|
; SI-NEXT: v_mov_b32_e32 v49, v18
|
|
; SI-NEXT: v_mov_b32_e32 v50, v17
|
|
; SI-NEXT: v_mov_b32_e32 v51, v16
|
|
; SI-NEXT: v_mov_b32_e32 v52, v15
|
|
; SI-NEXT: v_mov_b32_e32 v53, v14
|
|
; SI-NEXT: v_mov_b32_e32 v54, v13
|
|
; SI-NEXT: v_mov_b32_e32 v55, v12
|
|
; SI-NEXT: v_mov_b32_e32 v40, v11
|
|
; SI-NEXT: v_mov_b32_e32 v41, v10
|
|
; SI-NEXT: v_mov_b32_e32 v42, v9
|
|
; SI-NEXT: v_mov_b32_e32 v43, v8
|
|
; SI-NEXT: v_mov_b32_e32 v44, v7
|
|
; SI-NEXT: v_mov_b32_e32 v45, v6
|
|
; SI-NEXT: v_mov_b32_e32 v46, v5
|
|
; SI-NEXT: v_mov_b32_e32 v47, v4
|
|
; SI-NEXT: v_mov_b32_e32 v56, v3
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_mov_b32_e32 v57, v2
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_mov_b32_e32 v58, v1
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_mov_b32_e32 v59, v0
|
|
; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v26
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v25
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v23
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v22
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v21
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v8, 16, v48
|
|
; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v49
|
|
; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v50
|
|
; SI-NEXT: v_lshrrev_b32_e32 v11, 16, v51
|
|
; SI-NEXT: v_lshrrev_b32_e32 v12, 16, v52
|
|
; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v53
|
|
; SI-NEXT: v_lshrrev_b32_e32 v14, 16, v54
|
|
; SI-NEXT: v_lshrrev_b32_e32 v15, 16, v55
|
|
; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v40
|
|
; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v41
|
|
; SI-NEXT: v_lshrrev_b32_e32 v18, 16, v42
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v43
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v20, 16, v44
|
|
; SI-NEXT: v_lshrrev_b32_e32 v21, 16, v45
|
|
; SI-NEXT: v_lshrrev_b32_e32 v22, 16, v46
|
|
; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v47
|
|
; SI-NEXT: v_lshrrev_b32_e32 v24, 16, v56
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v57
|
|
; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v58
|
|
; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v59
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB42_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v55
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v59
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v58
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v57
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v56
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v47
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v46
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v45
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v44
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v43
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v42
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v41
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v39
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v38
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v62
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v37
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v32
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v36
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v61
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v35
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v60
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v34
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v63
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v33
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v54
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v53
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v52
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v51
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v50
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v49
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v27
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: .LBB42_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB42_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v55
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v59
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v58
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v57
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v56
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v47
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v46
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v45
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v44
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v43
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v42
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v41
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v40
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_or_b32_e32 v0, v39, v0
|
|
; SI-NEXT: s_mov_b32 s6, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v1, v38, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v62, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v37, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v32, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v36, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v61, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v35, v7
|
|
; SI-NEXT: v_or_b32_e32 v8, v60, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v34, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v63, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v33, v11
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v2
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, s6, v3
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v4
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, s6, v5
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v6
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, s6, v7
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v54
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v53
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v52
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v51
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v50
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v49
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v48
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v10
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, s6, v11
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v12
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, s6, v13
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v14
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, s6, v15
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v16
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, s6, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, s6, v19
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, s6, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, s6, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, s6, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 3, v25
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, s6, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 0x30000, v26
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 3, v27
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 0x30000, v27
|
|
; SI-NEXT: .LBB42_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56i16_to_v14i64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_mov_b32_e32 v32, v27
|
|
; VI-NEXT: v_mov_b32_e32 v33, v26
|
|
; VI-NEXT: v_mov_b32_e32 v34, v25
|
|
; VI-NEXT: v_mov_b32_e32 v35, v24
|
|
; VI-NEXT: v_mov_b32_e32 v36, v23
|
|
; VI-NEXT: v_mov_b32_e32 v37, v22
|
|
; VI-NEXT: v_mov_b32_e32 v38, v21
|
|
; VI-NEXT: v_mov_b32_e32 v39, v20
|
|
; VI-NEXT: v_mov_b32_e32 v48, v19
|
|
; VI-NEXT: v_mov_b32_e32 v49, v18
|
|
; VI-NEXT: v_mov_b32_e32 v50, v17
|
|
; VI-NEXT: v_mov_b32_e32 v51, v16
|
|
; VI-NEXT: v_mov_b32_e32 v52, v15
|
|
; VI-NEXT: v_mov_b32_e32 v53, v14
|
|
; VI-NEXT: v_mov_b32_e32 v54, v13
|
|
; VI-NEXT: v_mov_b32_e32 v55, v12
|
|
; VI-NEXT: v_mov_b32_e32 v40, v11
|
|
; VI-NEXT: v_mov_b32_e32 v41, v10
|
|
; VI-NEXT: v_mov_b32_e32 v42, v9
|
|
; VI-NEXT: v_mov_b32_e32 v43, v8
|
|
; VI-NEXT: v_mov_b32_e32 v44, v7
|
|
; VI-NEXT: v_mov_b32_e32 v45, v6
|
|
; VI-NEXT: v_mov_b32_e32 v46, v5
|
|
; VI-NEXT: v_mov_b32_e32 v47, v4
|
|
; VI-NEXT: v_mov_b32_e32 v56, v3
|
|
; VI-NEXT: v_mov_b32_e32 v57, v2
|
|
; VI-NEXT: v_mov_b32_e32 v58, v1
|
|
; VI-NEXT: v_mov_b32_e32 v59, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB42_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_mov_b32_e32 v27, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v27, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v27, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v27, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v27, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v27, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v27, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v27, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v27, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v27, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v27, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v27, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v27, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v27, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v27, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v27, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v27, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v27, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v27, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v27, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v27, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v27, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v24, v27, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v25, v27, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v26, v27, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v27, v27, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v56, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v47, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v46, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v45, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v44, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v43, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v42, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v41, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v40, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v53, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v52, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v51, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v50, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v49, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v39, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v37, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v36, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v35, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v34, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v33, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v32, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: ; implicit-def: $vgpr59
|
|
; VI-NEXT: ; implicit-def: $vgpr58
|
|
; VI-NEXT: ; implicit-def: $vgpr57
|
|
; VI-NEXT: ; implicit-def: $vgpr56
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: .LBB42_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB42_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 3
|
|
; VI-NEXT: v_add_u16_e32 v0, 3, v59
|
|
; VI-NEXT: v_add_u16_sdwa v1, v59, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v58
|
|
; VI-NEXT: v_add_u16_sdwa v3, v58, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v3
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v57
|
|
; VI-NEXT: v_add_u16_sdwa v3, v57, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; VI-NEXT: v_add_u16_e32 v3, 3, v56
|
|
; VI-NEXT: v_add_u16_sdwa v4, v56, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; VI-NEXT: v_add_u16_e32 v4, 3, v47
|
|
; VI-NEXT: v_add_u16_sdwa v5, v47, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; VI-NEXT: v_add_u16_e32 v5, 3, v46
|
|
; VI-NEXT: v_add_u16_sdwa v6, v46, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; VI-NEXT: v_add_u16_e32 v6, 3, v45
|
|
; VI-NEXT: v_add_u16_sdwa v7, v45, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; VI-NEXT: v_add_u16_e32 v7, 3, v44
|
|
; VI-NEXT: v_add_u16_sdwa v8, v44, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; VI-NEXT: v_add_u16_e32 v8, 3, v43
|
|
; VI-NEXT: v_add_u16_sdwa v9, v43, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; VI-NEXT: v_add_u16_e32 v9, 3, v42
|
|
; VI-NEXT: v_add_u16_sdwa v10, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; VI-NEXT: v_add_u16_e32 v10, 3, v41
|
|
; VI-NEXT: v_add_u16_sdwa v11, v41, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v10, v10, v11
|
|
; VI-NEXT: v_add_u16_e32 v11, 3, v40
|
|
; VI-NEXT: v_add_u16_sdwa v12, v40, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; VI-NEXT: v_add_u16_e32 v12, 3, v55
|
|
; VI-NEXT: v_add_u16_sdwa v13, v55, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; VI-NEXT: v_add_u16_e32 v13, 3, v54
|
|
; VI-NEXT: v_add_u16_sdwa v14, v54, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; VI-NEXT: v_add_u16_e32 v14, 3, v53
|
|
; VI-NEXT: v_add_u16_sdwa v15, v53, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; VI-NEXT: v_add_u16_e32 v15, 3, v52
|
|
; VI-NEXT: v_add_u16_sdwa v16, v52, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; VI-NEXT: v_add_u16_e32 v16, 3, v51
|
|
; VI-NEXT: v_add_u16_sdwa v17, v51, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; VI-NEXT: v_add_u16_e32 v17, 3, v50
|
|
; VI-NEXT: v_add_u16_sdwa v18, v50, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; VI-NEXT: v_add_u16_e32 v18, 3, v49
|
|
; VI-NEXT: v_add_u16_sdwa v19, v49, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; VI-NEXT: v_add_u16_e32 v19, 3, v48
|
|
; VI-NEXT: v_add_u16_sdwa v20, v48, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; VI-NEXT: v_add_u16_e32 v20, 3, v39
|
|
; VI-NEXT: v_add_u16_sdwa v21, v39, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; VI-NEXT: v_add_u16_e32 v21, 3, v38
|
|
; VI-NEXT: v_add_u16_sdwa v22, v38, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; VI-NEXT: v_add_u16_e32 v22, 3, v37
|
|
; VI-NEXT: v_add_u16_sdwa v23, v37, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; VI-NEXT: v_add_u16_e32 v23, 3, v36
|
|
; VI-NEXT: v_add_u16_sdwa v24, v36, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; VI-NEXT: v_add_u16_e32 v24, 3, v35
|
|
; VI-NEXT: v_add_u16_sdwa v25, v35, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; VI-NEXT: v_add_u16_e32 v25, 3, v34
|
|
; VI-NEXT: v_add_u16_sdwa v26, v34, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; VI-NEXT: v_add_u16_e32 v26, 3, v33
|
|
; VI-NEXT: v_add_u16_sdwa v28, v33, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; VI-NEXT: v_add_u16_e32 v28, 3, v32
|
|
; VI-NEXT: v_add_u16_sdwa v27, v32, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; VI-NEXT: .LBB42_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56i16_to_v14i64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_mov_b32_e32 v59, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v22
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v13
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v12
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v11
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v10
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v40
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v56, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v57, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v58, v1
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v41
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v47
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v56
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v57
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v58
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v59
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB42_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v37
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: .LBB42_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB42_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB42_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v14i64:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB42_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: .LBB42_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56i16_to_v14i64:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v69, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v71, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v80, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v68, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v67, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v66, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v65, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v64, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v54, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v53, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v52, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v51, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v50, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v49, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v48, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v39, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v38, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v37, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v36, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v35, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v34, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v33, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v32, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v31, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v30, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v29, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB42_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: .LBB42_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <56 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <56 x i16> %a1 to <14 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x i16> %a to <14 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x i64> %phi
|
|
}
|
|
|
|
define inreg <14 x i64> @bitcast_v56i16_to_v14i64_scalar(<56 x i16> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v56i16_to_v14i64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; SI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; SI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; SI-NEXT: v_writelane_b32 v28, s54, 14
|
|
; SI-NEXT: v_writelane_b32 v28, s55, 15
|
|
; SI-NEXT: v_writelane_b32 v28, s64, 16
|
|
; SI-NEXT: v_writelane_b32 v28, s65, 17
|
|
; SI-NEXT: v_writelane_b32 v28, s66, 18
|
|
; SI-NEXT: v_writelane_b32 v28, s67, 19
|
|
; SI-NEXT: v_writelane_b32 v28, s68, 20
|
|
; SI-NEXT: v_writelane_b32 v28, s69, 21
|
|
; SI-NEXT: v_writelane_b32 v28, s70, 22
|
|
; SI-NEXT: v_writelane_b32 v28, s71, 23
|
|
; SI-NEXT: v_writelane_b32 v28, s80, 24
|
|
; SI-NEXT: v_writelane_b32 v28, s81, 25
|
|
; SI-NEXT: v_writelane_b32 v28, s82, 26
|
|
; SI-NEXT: v_writelane_b32 v28, s83, 27
|
|
; SI-NEXT: v_writelane_b32 v28, s84, 28
|
|
; SI-NEXT: v_writelane_b32 v28, s85, 29
|
|
; SI-NEXT: v_writelane_b32 v28, s86, 30
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s73, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s75, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s77, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s88, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s94, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s31, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s69, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s80, v0
|
|
; SI-NEXT: v_writelane_b32 v28, s87, 31
|
|
; SI-NEXT: s_lshr_b32 s78, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s92, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s95, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s68, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s70, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s81, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s82, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s83, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s84, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s85, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s86, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s87, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s6, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s8, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s10, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s12, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s14, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s72, s73, 16
|
|
; SI-NEXT: s_lshr_b32 s74, s75, 16
|
|
; SI-NEXT: s_lshr_b32 s76, s77, 16
|
|
; SI-NEXT: s_lshr_b32 s79, s88, 16
|
|
; SI-NEXT: s_lshr_b32 s90, s91, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s94, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s31, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s69, 16
|
|
; SI-NEXT: s_lshr_b32 s71, s80, 16
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB43_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s36, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s37, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s38, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s39, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s40, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s41, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s42, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s43, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s44, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s34, 16
|
|
; SI-NEXT: s_or_b32 s45, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s95, 16
|
|
; SI-NEXT: s_or_b32 s46, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s92, 16
|
|
; SI-NEXT: s_or_b32 s47, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s48, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; SI-NEXT: s_or_b32 s49, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s80, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s50, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s51, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s31, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s52, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s94, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s53, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; SI-NEXT: s_or_b32 s54, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s55, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; SI-NEXT: s_or_b32 s56, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; SI-NEXT: s_or_b32 s57, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; SI-NEXT: s_or_b32 s58, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; SI-NEXT: s_or_b32 s59, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s12, 16
|
|
; SI-NEXT: s_or_b32 s60, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s10, 16
|
|
; SI-NEXT: s_or_b32 s61, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s8, 16
|
|
; SI-NEXT: s_or_b32 s62, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s6, 16
|
|
; SI-NEXT: s_or_b32 s63, s4, s5
|
|
; SI-NEXT: s_cbranch_execnz .LBB43_3
|
|
; SI-NEXT: .LBB43_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s17, s17, 3
|
|
; SI-NEXT: s_add_i32 s36, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: s_add_i32 s37, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s19, s19, 3
|
|
; SI-NEXT: s_add_i32 s38, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: s_add_i32 s39, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s21, s21, 3
|
|
; SI-NEXT: s_add_i32 s40, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: s_add_i32 s41, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s42, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s43, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s44, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s34, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s45, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s95, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s46, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s92, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s47, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: s_add_i32 s48, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s80, s80, 3
|
|
; SI-NEXT: s_add_i32 s49, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s80, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s69, s69, 3
|
|
; SI-NEXT: s_add_i32 s50, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s31, s31, 3
|
|
; SI-NEXT: s_add_i32 s51, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s31, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s94, s94, 3
|
|
; SI-NEXT: s_add_i32 s52, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s94, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s91, s91, 3
|
|
; SI-NEXT: s_add_i32 s53, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s88, s88, 3
|
|
; SI-NEXT: s_add_i32 s54, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s77, s77, 3
|
|
; SI-NEXT: s_add_i32 s55, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s75, s75, 3
|
|
; SI-NEXT: s_add_i32 s56, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s73, s73, 3
|
|
; SI-NEXT: s_add_i32 s57, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s15, s15, 3
|
|
; SI-NEXT: s_add_i32 s58, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s13, s13, 3
|
|
; SI-NEXT: s_add_i32 s59, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s12, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s11, s11, 3
|
|
; SI-NEXT: s_add_i32 s60, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s10, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s9, s9, 3
|
|
; SI-NEXT: s_add_i32 s61, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s8, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s7, s7, 3
|
|
; SI-NEXT: s_add_i32 s62, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s6, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s63, s4, 0x30000
|
|
; SI-NEXT: .LBB43_3: ; %end
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_readlane_b32 s87, v28, 31
|
|
; SI-NEXT: v_readlane_b32 s86, v28, 30
|
|
; SI-NEXT: v_readlane_b32 s85, v28, 29
|
|
; SI-NEXT: v_readlane_b32 s84, v28, 28
|
|
; SI-NEXT: v_readlane_b32 s83, v28, 27
|
|
; SI-NEXT: v_readlane_b32 s82, v28, 26
|
|
; SI-NEXT: v_readlane_b32 s81, v28, 25
|
|
; SI-NEXT: v_readlane_b32 s80, v28, 24
|
|
; SI-NEXT: v_readlane_b32 s71, v28, 23
|
|
; SI-NEXT: v_readlane_b32 s70, v28, 22
|
|
; SI-NEXT: v_readlane_b32 s69, v28, 21
|
|
; SI-NEXT: v_readlane_b32 s68, v28, 20
|
|
; SI-NEXT: v_readlane_b32 s67, v28, 19
|
|
; SI-NEXT: v_readlane_b32 s66, v28, 18
|
|
; SI-NEXT: v_readlane_b32 s65, v28, 17
|
|
; SI-NEXT: v_readlane_b32 s64, v28, 16
|
|
; SI-NEXT: v_readlane_b32 s55, v28, 15
|
|
; SI-NEXT: v_readlane_b32 s54, v28, 14
|
|
; SI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB43_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; SI-NEXT: s_branch .LBB43_2
|
|
;
|
|
; VI-LABEL: bitcast_v56i16_to_v14i64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; VI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; VI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; VI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; VI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; VI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; VI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; VI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; VI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; VI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; VI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; VI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; VI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; VI-NEXT: v_writelane_b32 v28, s54, 14
|
|
; VI-NEXT: v_writelane_b32 v28, s55, 15
|
|
; VI-NEXT: v_writelane_b32 v28, s64, 16
|
|
; VI-NEXT: v_writelane_b32 v28, s65, 17
|
|
; VI-NEXT: v_writelane_b32 v28, s66, 18
|
|
; VI-NEXT: v_writelane_b32 v28, s67, 19
|
|
; VI-NEXT: v_writelane_b32 v28, s68, 20
|
|
; VI-NEXT: v_writelane_b32 v28, s69, 21
|
|
; VI-NEXT: v_writelane_b32 v28, s70, 22
|
|
; VI-NEXT: v_writelane_b32 v28, s71, 23
|
|
; VI-NEXT: v_writelane_b32 v28, s80, 24
|
|
; VI-NEXT: v_writelane_b32 v28, s81, 25
|
|
; VI-NEXT: v_writelane_b32 v28, s82, 26
|
|
; VI-NEXT: v_writelane_b32 v28, s83, 27
|
|
; VI-NEXT: v_writelane_b32 v28, s84, 28
|
|
; VI-NEXT: v_writelane_b32 v28, s85, 29
|
|
; VI-NEXT: v_writelane_b32 v28, s86, 30
|
|
; VI-NEXT: v_writelane_b32 v28, s87, 31
|
|
; VI-NEXT: v_readfirstlane_b32 s86, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s87, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s73, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s77, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s88, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s34, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s69, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s81, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s84, v0
|
|
; VI-NEXT: s_lshr_b32 s79, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s67, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s70, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s80, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s82, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s85, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s7, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s11, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s13, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s15, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s10, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s64, s86, 16
|
|
; VI-NEXT: s_lshr_b32 s65, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s66, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s14, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s87, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s73, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s77, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s88, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s91, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s34, 16
|
|
; VI-NEXT: s_lshr_b32 s68, s69, 16
|
|
; VI-NEXT: s_lshr_b32 s71, s81, 16
|
|
; VI-NEXT: s_lshr_b32 s83, s84, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB43_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_or_b32 s36, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_or_b32 s38, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s10, 16
|
|
; VI-NEXT: s_or_b32 s39, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; VI-NEXT: s_or_b32 s37, s40, s41
|
|
; VI-NEXT: s_or_b32 s40, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_or_b32 s41, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s42, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s43, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; VI-NEXT: s_or_b32 s44, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s45, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s5, s67, 16
|
|
; VI-NEXT: s_or_b32 s46, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s47, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s48, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s49, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s84
|
|
; VI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; VI-NEXT: s_or_b32 s50, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s81
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s51, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s69
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s52, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s34
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s53, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s91
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s54, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s88
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s55, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s77
|
|
; VI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; VI-NEXT: s_or_b32 s56, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s73
|
|
; VI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; VI-NEXT: s_or_b32 s57, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; VI-NEXT: s_or_b32 s58, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s87
|
|
; VI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; VI-NEXT: s_or_b32 s59, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; VI-NEXT: s_or_b32 s60, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s5, s66, 16
|
|
; VI-NEXT: s_or_b32 s61, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s5, s65, 16
|
|
; VI-NEXT: s_or_b32 s62, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s86
|
|
; VI-NEXT: s_lshl_b32 s5, s64, 16
|
|
; VI-NEXT: s_or_b32 s63, s4, s5
|
|
; VI-NEXT: s_cbranch_execnz .LBB43_3
|
|
; VI-NEXT: .LBB43_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_and_b32 s5, s17, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s10, s10, 16
|
|
; VI-NEXT: s_or_b32 s5, s10, s5
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s36, s4, 0x30000
|
|
; VI-NEXT: s_add_i32 s37, s5, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s38, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s39, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s40, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s41, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s42, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s43, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s44, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s45, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s67, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s46, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s47, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s48, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s84, s84, 3
|
|
; VI-NEXT: s_add_i32 s49, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s84, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s81, s81, 3
|
|
; VI-NEXT: s_add_i32 s50, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s81, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s69, s69, 3
|
|
; VI-NEXT: s_add_i32 s51, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s34, s34, 3
|
|
; VI-NEXT: s_add_i32 s52, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s34, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s91, s91, 3
|
|
; VI-NEXT: s_add_i32 s53, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s88, s88, 3
|
|
; VI-NEXT: s_add_i32 s54, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s77, s77, 3
|
|
; VI-NEXT: s_add_i32 s55, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s75, s73, 3
|
|
; VI-NEXT: s_add_i32 s56, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s73, s12, 3
|
|
; VI-NEXT: s_add_i32 s57, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s15, s87, 3
|
|
; VI-NEXT: s_add_i32 s58, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s13, s8, 3
|
|
; VI-NEXT: s_add_i32 s59, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s11, s9, 3
|
|
; VI-NEXT: s_add_i32 s60, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s66, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s9, s6, 3
|
|
; VI-NEXT: s_add_i32 s61, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s65, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s7, s86, 3
|
|
; VI-NEXT: s_add_i32 s62, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s64, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s63, s4, 0x30000
|
|
; VI-NEXT: .LBB43_3: ; %end
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_readlane_b32 s87, v28, 31
|
|
; VI-NEXT: v_readlane_b32 s86, v28, 30
|
|
; VI-NEXT: v_readlane_b32 s85, v28, 29
|
|
; VI-NEXT: v_readlane_b32 s84, v28, 28
|
|
; VI-NEXT: v_readlane_b32 s83, v28, 27
|
|
; VI-NEXT: v_readlane_b32 s82, v28, 26
|
|
; VI-NEXT: v_readlane_b32 s81, v28, 25
|
|
; VI-NEXT: v_readlane_b32 s80, v28, 24
|
|
; VI-NEXT: v_readlane_b32 s71, v28, 23
|
|
; VI-NEXT: v_readlane_b32 s70, v28, 22
|
|
; VI-NEXT: v_readlane_b32 s69, v28, 21
|
|
; VI-NEXT: v_readlane_b32 s68, v28, 20
|
|
; VI-NEXT: v_readlane_b32 s67, v28, 19
|
|
; VI-NEXT: v_readlane_b32 s66, v28, 18
|
|
; VI-NEXT: v_readlane_b32 s65, v28, 17
|
|
; VI-NEXT: v_readlane_b32 s64, v28, 16
|
|
; VI-NEXT: v_readlane_b32 s55, v28, 15
|
|
; VI-NEXT: v_readlane_b32 s54, v28, 14
|
|
; VI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; VI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; VI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; VI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; VI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; VI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; VI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; VI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; VI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; VI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; VI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB43_4:
|
|
; VI-NEXT: ; implicit-def: $vgpr29 : SGPR spill to VGPR lane
|
|
; VI-NEXT: s_mov_b32 vcc_lo, s64
|
|
; VI-NEXT: v_writelane_b32 v29, s75, 0
|
|
; VI-NEXT: s_mov_b32 vcc_hi, s65
|
|
; VI-NEXT: s_mov_b32 s75, s73
|
|
; VI-NEXT: s_mov_b32 s73, s12
|
|
; VI-NEXT: v_writelane_b32 v29, s25, 1
|
|
; VI-NEXT: s_mov_b32 s25, s91
|
|
; VI-NEXT: s_mov_b32 s91, s26
|
|
; VI-NEXT: s_mov_b32 s26, s27
|
|
; VI-NEXT: s_mov_b32 s27, s88
|
|
; VI-NEXT: s_mov_b32 s88, s28
|
|
; VI-NEXT: s_mov_b32 s28, s77
|
|
; VI-NEXT: s_mov_b32 s77, s29
|
|
; VI-NEXT: s_mov_b32 s29, s15
|
|
; VI-NEXT: s_mov_b32 s15, s87
|
|
; VI-NEXT: s_mov_b32 s87, s85
|
|
; VI-NEXT: s_mov_b32 s85, s82
|
|
; VI-NEXT: s_mov_b32 s82, s71
|
|
; VI-NEXT: s_mov_b32 s71, s68
|
|
; VI-NEXT: s_mov_b32 s68, s35
|
|
; VI-NEXT: s_mov_b32 s35, s31
|
|
; VI-NEXT: s_mov_b32 s31, s30
|
|
; VI-NEXT: s_mov_b32 s30, s90
|
|
; VI-NEXT: s_mov_b32 s90, s89
|
|
; VI-NEXT: s_mov_b32 s89, s79
|
|
; VI-NEXT: s_mov_b32 s79, s78
|
|
; VI-NEXT: s_mov_b32 s78, s76
|
|
; VI-NEXT: s_mov_b32 s76, s74
|
|
; VI-NEXT: s_mov_b32 s74, s72
|
|
; VI-NEXT: s_mov_b32 s72, s14
|
|
; VI-NEXT: s_mov_b32 s14, s66
|
|
; VI-NEXT: s_mov_b32 s12, s10
|
|
; VI-NEXT: s_mov_b32 s10, s13
|
|
; VI-NEXT: s_mov_b32 s13, s8
|
|
; VI-NEXT: s_mov_b32 s8, s11
|
|
; VI-NEXT: s_mov_b32 s11, s9
|
|
; VI-NEXT: s_mov_b32 s9, s6
|
|
; VI-NEXT: s_mov_b32 s6, s7
|
|
; VI-NEXT: s_mov_b32 s7, s86
|
|
; VI-NEXT: s_mov_b32 s86, s83
|
|
; VI-NEXT: s_mov_b32 s83, s80
|
|
; VI-NEXT: s_mov_b32 s80, s70
|
|
; VI-NEXT: s_mov_b32 s70, s67
|
|
; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; VI-NEXT: s_mov_b32 s67, s70
|
|
; VI-NEXT: s_mov_b32 s70, s80
|
|
; VI-NEXT: s_mov_b32 s80, s83
|
|
; VI-NEXT: s_mov_b32 s83, s86
|
|
; VI-NEXT: s_mov_b32 s86, s7
|
|
; VI-NEXT: s_mov_b32 s7, s6
|
|
; VI-NEXT: s_mov_b32 s6, s9
|
|
; VI-NEXT: s_mov_b32 s9, s11
|
|
; VI-NEXT: s_mov_b32 s11, s8
|
|
; VI-NEXT: s_mov_b32 s8, s13
|
|
; VI-NEXT: s_mov_b32 s13, s10
|
|
; VI-NEXT: s_mov_b32 s10, s12
|
|
; VI-NEXT: s_mov_b32 s66, s14
|
|
; VI-NEXT: s_mov_b32 s14, s72
|
|
; VI-NEXT: s_mov_b32 s72, s74
|
|
; VI-NEXT: s_mov_b32 s74, s76
|
|
; VI-NEXT: s_mov_b32 s76, s78
|
|
; VI-NEXT: s_mov_b32 s78, s79
|
|
; VI-NEXT: s_mov_b32 s79, s89
|
|
; VI-NEXT: s_mov_b32 s89, s90
|
|
; VI-NEXT: s_mov_b32 s90, s30
|
|
; VI-NEXT: s_mov_b32 s30, s31
|
|
; VI-NEXT: s_mov_b32 s31, s35
|
|
; VI-NEXT: s_mov_b32 s35, s68
|
|
; VI-NEXT: s_mov_b32 s68, s71
|
|
; VI-NEXT: s_mov_b32 s71, s82
|
|
; VI-NEXT: s_mov_b32 s82, s85
|
|
; VI-NEXT: s_mov_b32 s85, s87
|
|
; VI-NEXT: s_mov_b32 s87, s15
|
|
; VI-NEXT: s_mov_b32 s15, s29
|
|
; VI-NEXT: s_mov_b32 s29, s77
|
|
; VI-NEXT: s_mov_b32 s77, s28
|
|
; VI-NEXT: s_mov_b32 s28, s88
|
|
; VI-NEXT: s_mov_b32 s88, s27
|
|
; VI-NEXT: s_mov_b32 s27, s26
|
|
; VI-NEXT: s_mov_b32 s26, s91
|
|
; VI-NEXT: s_mov_b32 s91, s25
|
|
; VI-NEXT: v_readlane_b32 s25, v29, 1
|
|
; VI-NEXT: s_mov_b32 s12, s73
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: v_readlane_b32 s75, v29, 0
|
|
; VI-NEXT: s_mov_b32 s65, vcc_hi
|
|
; VI-NEXT: s_mov_b32 s64, vcc_lo
|
|
; VI-NEXT: s_branch .LBB43_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v56i16_to_v14i64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: s_lshr_b32 s12, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s13, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s14, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s15, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s88, v5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v3
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v2
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: s_lshr_b32 s4, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s5, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s6, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s7, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s8, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s9, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s10, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s63, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s62, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s61, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s60, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s59, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s58, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s57, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s56, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s88, 16
|
|
; GFX9-NEXT: s_lshr_b32 s17, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s18, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s19, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s20, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s21, s12, 16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_cmp_lg_u32 s42, 0
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s50, s12, s21
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s51, s13, s20
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s52, s14, s19
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s53, s15, s18
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s54, s16, s17
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s55, s88, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s56, s56, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s57, s57, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s58, s58, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s59, s59, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s60, s60, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s61, s61, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s62, s62, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s63, s63, s72
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB43_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB43_4
|
|
; GFX9-NEXT: .LBB43_2: ; %cmp.true
|
|
; GFX9-NEXT: v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_branch .LBB43_5
|
|
; GFX9-NEXT: .LBB43_3:
|
|
; GFX9-NEXT: s_branch .LBB43_2
|
|
; GFX9-NEXT: .LBB43_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB43_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v56i16_to_v14i64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s44, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s74, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s76, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s77, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s78, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s88, v10
|
|
; GFX11-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s0, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s44, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s56, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s58, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s59, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s60, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s62, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s74, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s76, 16
|
|
; GFX11-NEXT: s_lshr_b32 s90, s77, 16
|
|
; GFX11-NEXT: s_lshr_b32 s91, s78, 16
|
|
; GFX11-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s88, 0
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s78, s91
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s77, s90
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s76, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s74, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s62, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s60, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s59, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s58, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s56, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s44, s57
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB43_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB43_4
|
|
; GFX11-NEXT: .LBB43_2: ; %cmp.true
|
|
; GFX11-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v18, s18, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v19, s19, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v20, s20, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v21, s21, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v22, s22, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v23, s23, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v24, s24, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v25, s25, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v26, s26, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v27, s27, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB43_3:
|
|
; GFX11-NEXT: s_branch .LBB43_2
|
|
; GFX11-NEXT: .LBB43_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <56 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <56 x i16> %a1 to <14 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x i16> %a to <14 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x i64> %phi
|
|
}
|
|
|
|
define <56 x half> @bitcast_v14i64_to_v56f16(<14 x i64> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v14i64_to_v56f16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB44_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v37, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v39, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB44_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB44_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v37, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v39, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB44_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v47
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v40
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v41
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v53
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v50
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v39
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v37
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14i64_to_v56f16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB44_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB44_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB44_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_add_u32_e32 v26, vcc, 3, v26
|
|
; VI-NEXT: v_addc_u32_e32 v27, vcc, 0, v27, vcc
|
|
; VI-NEXT: v_add_u32_e32 v24, vcc, 3, v24
|
|
; VI-NEXT: v_addc_u32_e32 v25, vcc, 0, v25, vcc
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 3, v22
|
|
; VI-NEXT: v_addc_u32_e32 v23, vcc, 0, v23, vcc
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_addc_u32_e32 v21, vcc, 0, v21, vcc
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_addc_u32_e32 v19, vcc, 0, v19, vcc
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_addc_u32_e32 v17, vcc, 0, v17, vcc
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 3, v14
|
|
; VI-NEXT: v_addc_u32_e32 v15, vcc, 0, v15, vcc
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_addc_u32_e32 v11, vcc, 0, v11, vcc
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_addc_u32_e32 v9, vcc, 0, v9, vcc
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB44_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14i64_to_v56f16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB44_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB44_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB44_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: v_add_co_u32_e32 v26, vcc, 3, v26
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v24, vcc, 3, v24
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v22, vcc, 3, v22
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v20, vcc, 3, v20
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v18, vcc, 3, v18
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v16, vcc, 3, v16
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v14, vcc, 3, v14
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v12, vcc, 3, v12
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v10, vcc, 3, v10
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v8, vcc, 3, v8
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v6, vcc, 3, v6
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v4, vcc, 3, v4
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v2, vcc, 3, v2
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
|
|
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 3, v0
|
|
; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB44_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v47, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v46, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v45, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v44, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v43, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v42, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v41, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v40, v7, s4
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v8, v55, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v54, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v53, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v52, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v51, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v49, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v48, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v39, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v38, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v37, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v36, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v35, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v34, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v33, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v32, v23, s4
|
|
; GFX9-NEXT: v_perm_b32 v24, v31, v24, s4
|
|
; GFX9-NEXT: v_perm_b32 v25, v30, v25, s4
|
|
; GFX9-NEXT: v_perm_b32 v26, v29, v26, s4
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s4
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v14i64_to_v56f16:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB44_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
|
|
; GFX11-TRUE16-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
|
|
; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
|
|
; GFX11-TRUE16-NEXT: .LBB44_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v14i64_to_v56f16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr67
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB44_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB44_2: ; %Flow
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB44_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v26, vcc_lo, v26, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v24, vcc_lo, v24, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v22, vcc_lo, v22, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v20, vcc_lo, v20, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v18, vcc_lo, v18, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v16, vcc_lo, v16, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v14, vcc_lo, v14, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v12, vcc_lo, v12, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v10, vcc_lo, v10, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v8, vcc_lo, v8, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v6, vcc_lo, v6, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v4, vcc_lo, v4, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v2, vcc_lo, v2, 3
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_add_co_u32 v0, vcc_lo, v0, 3
|
|
; GFX11-FAKE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB44_4: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v71, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v69, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v68, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v67, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v66, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v65, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v64, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v55, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v54, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v53, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v52, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v51, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v49, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v39, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v37, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v36, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v35, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v34, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v33, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v32, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v31, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v30, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v29, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <14 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <14 x i64> %a1 to <56 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x i64> %a to <56 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x half> %phi
|
|
}
|
|
|
|
define inreg <56 x half> @bitcast_v14i64_to_v56f16_scalar(<14 x i64> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v14i64_to_v56f16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v14
|
|
; SI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; SI-NEXT: v_readfirstlane_b32 s5, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s40, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v0
|
|
; SI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; SI-NEXT: s_cbranch_scc0 .LBB45_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s30, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s53, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB45_3
|
|
; SI-NEXT: .LBB45_2: ; %cmp.true
|
|
; SI-NEXT: s_add_u32 s4, s4, 3
|
|
; SI-NEXT: s_addc_u32 s5, s5, 0
|
|
; SI-NEXT: s_add_u32 s6, s6, 3
|
|
; SI-NEXT: s_addc_u32 s7, s7, 0
|
|
; SI-NEXT: s_add_u32 s8, s8, 3
|
|
; SI-NEXT: s_addc_u32 s9, s9, 0
|
|
; SI-NEXT: s_add_u32 s10, s10, 3
|
|
; SI-NEXT: s_addc_u32 s11, s11, 0
|
|
; SI-NEXT: s_add_u32 s12, s12, 3
|
|
; SI-NEXT: s_addc_u32 s13, s13, 0
|
|
; SI-NEXT: s_add_u32 s14, s14, 3
|
|
; SI-NEXT: s_addc_u32 s15, s15, 0
|
|
; SI-NEXT: s_add_u32 s40, s40, 3
|
|
; SI-NEXT: s_addc_u32 s41, s41, 0
|
|
; SI-NEXT: s_add_u32 s28, s28, 3
|
|
; SI-NEXT: s_addc_u32 s29, s29, 0
|
|
; SI-NEXT: s_add_u32 s26, s26, 3
|
|
; SI-NEXT: s_addc_u32 s27, s27, 0
|
|
; SI-NEXT: s_add_u32 s24, s24, 3
|
|
; SI-NEXT: s_addc_u32 s25, s25, 0
|
|
; SI-NEXT: s_add_u32 s22, s22, 3
|
|
; SI-NEXT: s_addc_u32 s23, s23, 0
|
|
; SI-NEXT: s_add_u32 s20, s20, 3
|
|
; SI-NEXT: s_addc_u32 s21, s21, 0
|
|
; SI-NEXT: s_add_u32 s18, s18, 3
|
|
; SI-NEXT: s_addc_u32 s19, s19, 0
|
|
; SI-NEXT: s_add_u32 s16, s16, 3
|
|
; SI-NEXT: s_addc_u32 s17, s17, 0
|
|
; SI-NEXT: s_lshr_b32 s30, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s53, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: .LBB45_3: ; %end
|
|
; SI-NEXT: s_lshl_b32 s43, s92, 16
|
|
; SI-NEXT: s_and_b32 s16, s16, 0xffff
|
|
; SI-NEXT: s_or_b32 s16, s16, s43
|
|
; SI-NEXT: s_and_b32 s17, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s53, 16
|
|
; SI-NEXT: s_or_b32 s17, s17, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s90, 16
|
|
; SI-NEXT: s_and_b32 s18, s18, 0xffff
|
|
; SI-NEXT: s_or_b32 s18, s18, s43
|
|
; SI-NEXT: s_and_b32 s19, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s52, 16
|
|
; SI-NEXT: s_or_b32 s19, s19, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s88, 16
|
|
; SI-NEXT: s_and_b32 s20, s20, 0xffff
|
|
; SI-NEXT: s_or_b32 s20, s20, s43
|
|
; SI-NEXT: s_and_b32 s21, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s51, 16
|
|
; SI-NEXT: s_or_b32 s21, s21, s43
|
|
; SI-NEXT: s_lshl_b32 s43, s78, 16
|
|
; SI-NEXT: s_and_b32 s22, s22, 0xffff
|
|
; SI-NEXT: s_or_b32 s22, s22, s43
|
|
; SI-NEXT: s_and_b32 s23, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s50, 16
|
|
; SI-NEXT: s_or_b32 s23, s23, s43
|
|
; SI-NEXT: s_and_b32 s24, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s76, 16
|
|
; SI-NEXT: s_or_b32 s24, s24, s43
|
|
; SI-NEXT: s_and_b32 s25, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s49, 16
|
|
; SI-NEXT: s_or_b32 s25, s25, s43
|
|
; SI-NEXT: s_and_b32 s26, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s74, 16
|
|
; SI-NEXT: s_or_b32 s26, s26, s43
|
|
; SI-NEXT: s_and_b32 s27, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s48, 16
|
|
; SI-NEXT: s_or_b32 s27, s27, s43
|
|
; SI-NEXT: s_and_b32 s28, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s72, 16
|
|
; SI-NEXT: s_or_b32 s28, s28, s43
|
|
; SI-NEXT: s_and_b32 s29, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s39, 16
|
|
; SI-NEXT: s_or_b32 s29, s29, s43
|
|
; SI-NEXT: s_and_b32 s40, s40, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s62, 16
|
|
; SI-NEXT: s_or_b32 s40, s40, s43
|
|
; SI-NEXT: s_and_b32 s41, s41, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s38, 16
|
|
; SI-NEXT: s_or_b32 s41, s41, s43
|
|
; SI-NEXT: s_and_b32 s14, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s60, 16
|
|
; SI-NEXT: s_or_b32 s14, s14, s43
|
|
; SI-NEXT: s_and_b32 s15, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s37, 16
|
|
; SI-NEXT: s_or_b32 s15, s15, s43
|
|
; SI-NEXT: s_and_b32 s12, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s58, 16
|
|
; SI-NEXT: s_or_b32 s12, s12, s43
|
|
; SI-NEXT: s_and_b32 s13, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s36, 16
|
|
; SI-NEXT: s_or_b32 s13, s13, s43
|
|
; SI-NEXT: s_and_b32 s10, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s56, 16
|
|
; SI-NEXT: s_or_b32 s10, s10, s43
|
|
; SI-NEXT: s_and_b32 s11, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s35, 16
|
|
; SI-NEXT: s_or_b32 s11, s11, s43
|
|
; SI-NEXT: s_and_b32 s8, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s46, 16
|
|
; SI-NEXT: s_or_b32 s8, s8, s43
|
|
; SI-NEXT: s_and_b32 s9, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s34, 16
|
|
; SI-NEXT: s_or_b32 s9, s9, s43
|
|
; SI-NEXT: s_and_b32 s6, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s44, 16
|
|
; SI-NEXT: s_and_b32 s4, s4, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s42, 16
|
|
; SI-NEXT: s_or_b32 s6, s6, s43
|
|
; SI-NEXT: s_and_b32 s7, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s43, s31, 16
|
|
; SI-NEXT: s_or_b32 s4, s4, s42
|
|
; SI-NEXT: s_and_b32 s5, s5, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s30, 16
|
|
; SI-NEXT: s_or_b32 s7, s7, s43
|
|
; SI-NEXT: s_or_b32 s5, s5, s42
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v14, s40
|
|
; SI-NEXT: v_mov_b32_e32 v15, s41
|
|
; SI-NEXT: v_mov_b32_e32 v16, s14
|
|
; SI-NEXT: v_mov_b32_e32 v17, s15
|
|
; SI-NEXT: v_mov_b32_e32 v18, s12
|
|
; SI-NEXT: v_mov_b32_e32 v19, s13
|
|
; SI-NEXT: v_mov_b32_e32 v20, s10
|
|
; SI-NEXT: v_mov_b32_e32 v21, s11
|
|
; SI-NEXT: v_mov_b32_e32 v22, s8
|
|
; SI-NEXT: v_mov_b32_e32 v23, s9
|
|
; SI-NEXT: v_mov_b32_e32 v24, s6
|
|
; SI-NEXT: v_mov_b32_e32 v25, s7
|
|
; SI-NEXT: v_mov_b32_e32 v26, s4
|
|
; SI-NEXT: v_mov_b32_e32 v27, s5
|
|
; SI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB45_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr92
|
|
; SI-NEXT: ; implicit-def: $sgpr53
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: ; implicit-def: $sgpr52
|
|
; SI-NEXT: ; implicit-def: $sgpr88
|
|
; SI-NEXT: ; implicit-def: $sgpr51
|
|
; SI-NEXT: ; implicit-def: $sgpr78
|
|
; SI-NEXT: ; implicit-def: $sgpr50
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr49
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr48
|
|
; SI-NEXT: ; implicit-def: $sgpr72
|
|
; SI-NEXT: ; implicit-def: $sgpr39
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr38
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr37
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr36
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr35
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr31
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: s_branch .LBB45_2
|
|
;
|
|
; VI-LABEL: bitcast_v14i64_to_v56f16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v0
|
|
; VI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB45_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s44, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s42, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB45_3
|
|
; VI-NEXT: .LBB45_2: ; %cmp.true
|
|
; VI-NEXT: s_add_u32 s7, s7, 3
|
|
; VI-NEXT: s_addc_u32 s6, s6, 0
|
|
; VI-NEXT: s_add_u32 s9, s9, 3
|
|
; VI-NEXT: s_addc_u32 s8, s8, 0
|
|
; VI-NEXT: s_add_u32 s11, s11, 3
|
|
; VI-NEXT: s_addc_u32 s10, s10, 0
|
|
; VI-NEXT: s_add_u32 s13, s13, 3
|
|
; VI-NEXT: s_addc_u32 s12, s12, 0
|
|
; VI-NEXT: s_add_u32 s15, s15, 3
|
|
; VI-NEXT: s_addc_u32 s14, s14, 0
|
|
; VI-NEXT: s_add_u32 s41, s41, 3
|
|
; VI-NEXT: s_addc_u32 s40, s40, 0
|
|
; VI-NEXT: s_add_u32 s43, s43, 3
|
|
; VI-NEXT: s_addc_u32 s42, s42, 0
|
|
; VI-NEXT: s_add_u32 s28, s28, 3
|
|
; VI-NEXT: s_addc_u32 s29, s29, 0
|
|
; VI-NEXT: s_add_u32 s26, s26, 3
|
|
; VI-NEXT: s_addc_u32 s27, s27, 0
|
|
; VI-NEXT: s_add_u32 s24, s24, 3
|
|
; VI-NEXT: s_addc_u32 s25, s25, 0
|
|
; VI-NEXT: s_add_u32 s22, s22, 3
|
|
; VI-NEXT: s_addc_u32 s23, s23, 0
|
|
; VI-NEXT: s_add_u32 s20, s20, 3
|
|
; VI-NEXT: s_addc_u32 s21, s21, 0
|
|
; VI-NEXT: s_add_u32 s18, s18, 3
|
|
; VI-NEXT: s_addc_u32 s19, s19, 0
|
|
; VI-NEXT: s_add_u32 s16, s16, 3
|
|
; VI-NEXT: s_addc_u32 s17, s17, 0
|
|
; VI-NEXT: s_lshr_b32 s44, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s42, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: .LBB45_3: ; %end
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s16, s34, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s16
|
|
; VI-NEXT: s_and_b32 s16, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s17, s31, 16
|
|
; VI-NEXT: s_or_b32 s16, s16, s17
|
|
; VI-NEXT: s_and_b32 s17, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s18, s30, 16
|
|
; VI-NEXT: s_or_b32 s17, s17, s18
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s19, s91, 16
|
|
; VI-NEXT: s_or_b32 s18, s18, s19
|
|
; VI-NEXT: s_and_b32 s19, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s20, s90, 16
|
|
; VI-NEXT: s_or_b32 s19, s19, s20
|
|
; VI-NEXT: s_and_b32 s20, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s21, s89, 16
|
|
; VI-NEXT: s_or_b32 s20, s20, s21
|
|
; VI-NEXT: s_and_b32 s21, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s22, s88, 16
|
|
; VI-NEXT: s_or_b32 s21, s21, s22
|
|
; VI-NEXT: s_and_b32 s22, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s23, s79, 16
|
|
; VI-NEXT: s_or_b32 s22, s22, s23
|
|
; VI-NEXT: s_and_b32 s23, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s24, s78, 16
|
|
; VI-NEXT: s_or_b32 s23, s23, s24
|
|
; VI-NEXT: s_and_b32 s24, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s25, s77, 16
|
|
; VI-NEXT: s_or_b32 s24, s24, s25
|
|
; VI-NEXT: s_and_b32 s25, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s26, s76, 16
|
|
; VI-NEXT: s_or_b32 s25, s25, s26
|
|
; VI-NEXT: s_and_b32 s26, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s27, s75, 16
|
|
; VI-NEXT: s_or_b32 s26, s26, s27
|
|
; VI-NEXT: s_and_b32 s27, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s28, s74, 16
|
|
; VI-NEXT: s_or_b32 s27, s27, s28
|
|
; VI-NEXT: s_and_b32 s28, 0xffff, s43
|
|
; VI-NEXT: s_lshl_b32 s29, s73, 16
|
|
; VI-NEXT: s_or_b32 s28, s28, s29
|
|
; VI-NEXT: s_and_b32 s29, 0xffff, s42
|
|
; VI-NEXT: s_lshl_b32 s42, s72, 16
|
|
; VI-NEXT: s_or_b32 s29, s29, s42
|
|
; VI-NEXT: s_and_b32 s41, 0xffff, s41
|
|
; VI-NEXT: s_lshl_b32 s42, s63, 16
|
|
; VI-NEXT: s_or_b32 s41, s41, s42
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s40
|
|
; VI-NEXT: s_lshl_b32 s42, s62, 16
|
|
; VI-NEXT: s_or_b32 s40, s40, s42
|
|
; VI-NEXT: s_and_b32 s15, 0xffff, s15
|
|
; VI-NEXT: s_lshl_b32 s42, s61, 16
|
|
; VI-NEXT: s_or_b32 s15, s15, s42
|
|
; VI-NEXT: s_and_b32 s14, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s42, s60, 16
|
|
; VI-NEXT: s_or_b32 s14, s14, s42
|
|
; VI-NEXT: s_and_b32 s13, 0xffff, s13
|
|
; VI-NEXT: s_lshl_b32 s42, s59, 16
|
|
; VI-NEXT: s_or_b32 s13, s13, s42
|
|
; VI-NEXT: s_and_b32 s12, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s42, s58, 16
|
|
; VI-NEXT: s_or_b32 s12, s12, s42
|
|
; VI-NEXT: s_and_b32 s11, 0xffff, s11
|
|
; VI-NEXT: s_lshl_b32 s42, s57, 16
|
|
; VI-NEXT: s_or_b32 s11, s11, s42
|
|
; VI-NEXT: s_and_b32 s10, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s42, s56, 16
|
|
; VI-NEXT: s_or_b32 s10, s10, s42
|
|
; VI-NEXT: s_and_b32 s9, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s42, s47, 16
|
|
; VI-NEXT: s_or_b32 s9, s9, s42
|
|
; VI-NEXT: s_and_b32 s8, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s42, s46, 16
|
|
; VI-NEXT: s_or_b32 s8, s8, s42
|
|
; VI-NEXT: s_and_b32 s7, 0xffff, s7
|
|
; VI-NEXT: s_lshl_b32 s42, s45, 16
|
|
; VI-NEXT: s_or_b32 s7, s7, s42
|
|
; VI-NEXT: s_and_b32 s6, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s42, s44, 16
|
|
; VI-NEXT: s_or_b32 s6, s6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s16
|
|
; VI-NEXT: v_mov_b32_e32 v3, s17
|
|
; VI-NEXT: v_mov_b32_e32 v4, s18
|
|
; VI-NEXT: v_mov_b32_e32 v5, s19
|
|
; VI-NEXT: v_mov_b32_e32 v6, s20
|
|
; VI-NEXT: v_mov_b32_e32 v7, s21
|
|
; VI-NEXT: v_mov_b32_e32 v8, s22
|
|
; VI-NEXT: v_mov_b32_e32 v9, s23
|
|
; VI-NEXT: v_mov_b32_e32 v10, s24
|
|
; VI-NEXT: v_mov_b32_e32 v11, s25
|
|
; VI-NEXT: v_mov_b32_e32 v12, s26
|
|
; VI-NEXT: v_mov_b32_e32 v13, s27
|
|
; VI-NEXT: v_mov_b32_e32 v14, s28
|
|
; VI-NEXT: v_mov_b32_e32 v15, s29
|
|
; VI-NEXT: v_mov_b32_e32 v16, s41
|
|
; VI-NEXT: v_mov_b32_e32 v17, s40
|
|
; VI-NEXT: v_mov_b32_e32 v18, s15
|
|
; VI-NEXT: v_mov_b32_e32 v19, s14
|
|
; VI-NEXT: v_mov_b32_e32 v20, s13
|
|
; VI-NEXT: v_mov_b32_e32 v21, s12
|
|
; VI-NEXT: v_mov_b32_e32 v22, s11
|
|
; VI-NEXT: v_mov_b32_e32 v23, s10
|
|
; VI-NEXT: v_mov_b32_e32 v24, s9
|
|
; VI-NEXT: v_mov_b32_e32 v25, s8
|
|
; VI-NEXT: v_mov_b32_e32 v26, s7
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB45_4:
|
|
; VI-NEXT: ; implicit-def: $sgpr35
|
|
; VI-NEXT: ; implicit-def: $sgpr34
|
|
; VI-NEXT: ; implicit-def: $sgpr31
|
|
; VI-NEXT: ; implicit-def: $sgpr30
|
|
; VI-NEXT: ; implicit-def: $sgpr91
|
|
; VI-NEXT: ; implicit-def: $sgpr90
|
|
; VI-NEXT: ; implicit-def: $sgpr89
|
|
; VI-NEXT: ; implicit-def: $sgpr88
|
|
; VI-NEXT: ; implicit-def: $sgpr79
|
|
; VI-NEXT: ; implicit-def: $sgpr78
|
|
; VI-NEXT: ; implicit-def: $sgpr77
|
|
; VI-NEXT: ; implicit-def: $sgpr76
|
|
; VI-NEXT: ; implicit-def: $sgpr75
|
|
; VI-NEXT: ; implicit-def: $sgpr74
|
|
; VI-NEXT: ; implicit-def: $sgpr73
|
|
; VI-NEXT: ; implicit-def: $sgpr72
|
|
; VI-NEXT: ; implicit-def: $sgpr63
|
|
; VI-NEXT: ; implicit-def: $sgpr62
|
|
; VI-NEXT: ; implicit-def: $sgpr61
|
|
; VI-NEXT: ; implicit-def: $sgpr60
|
|
; VI-NEXT: ; implicit-def: $sgpr59
|
|
; VI-NEXT: ; implicit-def: $sgpr58
|
|
; VI-NEXT: ; implicit-def: $sgpr57
|
|
; VI-NEXT: ; implicit-def: $sgpr56
|
|
; VI-NEXT: ; implicit-def: $sgpr47
|
|
; VI-NEXT: ; implicit-def: $sgpr46
|
|
; VI-NEXT: ; implicit-def: $sgpr45
|
|
; VI-NEXT: ; implicit-def: $sgpr44
|
|
; VI-NEXT: s_branch .LBB45_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v14i64_to_v56f16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s43, v0
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB45_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s44, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s42, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s43, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB45_3
|
|
; GFX9-NEXT: .LBB45_2: ; %cmp.true
|
|
; GFX9-NEXT: s_add_u32 s7, s7, 3
|
|
; GFX9-NEXT: s_addc_u32 s6, s6, 0
|
|
; GFX9-NEXT: s_add_u32 s9, s9, 3
|
|
; GFX9-NEXT: s_addc_u32 s8, s8, 0
|
|
; GFX9-NEXT: s_add_u32 s11, s11, 3
|
|
; GFX9-NEXT: s_addc_u32 s10, s10, 0
|
|
; GFX9-NEXT: s_add_u32 s13, s13, 3
|
|
; GFX9-NEXT: s_addc_u32 s12, s12, 0
|
|
; GFX9-NEXT: s_add_u32 s15, s15, 3
|
|
; GFX9-NEXT: s_addc_u32 s14, s14, 0
|
|
; GFX9-NEXT: s_add_u32 s41, s41, 3
|
|
; GFX9-NEXT: s_addc_u32 s40, s40, 0
|
|
; GFX9-NEXT: s_add_u32 s43, s43, 3
|
|
; GFX9-NEXT: s_addc_u32 s42, s42, 0
|
|
; GFX9-NEXT: s_add_u32 s28, s28, 3
|
|
; GFX9-NEXT: s_addc_u32 s29, s29, 0
|
|
; GFX9-NEXT: s_add_u32 s26, s26, 3
|
|
; GFX9-NEXT: s_addc_u32 s27, s27, 0
|
|
; GFX9-NEXT: s_add_u32 s24, s24, 3
|
|
; GFX9-NEXT: s_addc_u32 s25, s25, 0
|
|
; GFX9-NEXT: s_add_u32 s22, s22, 3
|
|
; GFX9-NEXT: s_addc_u32 s23, s23, 0
|
|
; GFX9-NEXT: s_add_u32 s20, s20, 3
|
|
; GFX9-NEXT: s_addc_u32 s21, s21, 0
|
|
; GFX9-NEXT: s_add_u32 s18, s18, 3
|
|
; GFX9-NEXT: s_addc_u32 s19, s19, 0
|
|
; GFX9-NEXT: s_add_u32 s16, s16, 3
|
|
; GFX9-NEXT: s_addc_u32 s17, s17, 0
|
|
; GFX9-NEXT: s_lshr_b32 s44, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s42, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s43, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: .LBB45_3: ; %end
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s95
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s94
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s93
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s92
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s91
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s90
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s88
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s28, s43, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s29, s42, s72
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s41, s63
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s40, s62
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s15, s61
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s14, s60
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s13, s59
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s12, s58
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s57
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s56
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s47
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s46
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s45
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s4
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s5
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s6
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB45_4:
|
|
; GFX9-NEXT: ; implicit-def: $sgpr95
|
|
; GFX9-NEXT: ; implicit-def: $sgpr94
|
|
; GFX9-NEXT: ; implicit-def: $sgpr93
|
|
; GFX9-NEXT: ; implicit-def: $sgpr92
|
|
; GFX9-NEXT: ; implicit-def: $sgpr91
|
|
; GFX9-NEXT: ; implicit-def: $sgpr90
|
|
; GFX9-NEXT: ; implicit-def: $sgpr89
|
|
; GFX9-NEXT: ; implicit-def: $sgpr88
|
|
; GFX9-NEXT: ; implicit-def: $sgpr79
|
|
; GFX9-NEXT: ; implicit-def: $sgpr78
|
|
; GFX9-NEXT: ; implicit-def: $sgpr77
|
|
; GFX9-NEXT: ; implicit-def: $sgpr76
|
|
; GFX9-NEXT: ; implicit-def: $sgpr75
|
|
; GFX9-NEXT: ; implicit-def: $sgpr74
|
|
; GFX9-NEXT: ; implicit-def: $sgpr73
|
|
; GFX9-NEXT: ; implicit-def: $sgpr72
|
|
; GFX9-NEXT: ; implicit-def: $sgpr63
|
|
; GFX9-NEXT: ; implicit-def: $sgpr62
|
|
; GFX9-NEXT: ; implicit-def: $sgpr61
|
|
; GFX9-NEXT: ; implicit-def: $sgpr60
|
|
; GFX9-NEXT: ; implicit-def: $sgpr59
|
|
; GFX9-NEXT: ; implicit-def: $sgpr58
|
|
; GFX9-NEXT: ; implicit-def: $sgpr57
|
|
; GFX9-NEXT: ; implicit-def: $sgpr56
|
|
; GFX9-NEXT: ; implicit-def: $sgpr47
|
|
; GFX9-NEXT: ; implicit-def: $sgpr46
|
|
; GFX9-NEXT: ; implicit-def: $sgpr45
|
|
; GFX9-NEXT: ; implicit-def: $sgpr44
|
|
; GFX9-NEXT: s_branch .LBB45_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v14i64_to_v56f16_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-NEXT: s_mov_b32 s90, 0
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB45_4
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-NEXT: s_lshr_b32 s14, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s74, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s76, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s77, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s78, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s88, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s0, 16
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s90
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB45_3
|
|
; GFX11-NEXT: .LBB45_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_u32 s5, s5, 3
|
|
; GFX11-NEXT: s_addc_u32 s4, s4, 0
|
|
; GFX11-NEXT: s_add_u32 s7, s7, 3
|
|
; GFX11-NEXT: s_addc_u32 s6, s6, 0
|
|
; GFX11-NEXT: s_add_u32 s9, s9, 3
|
|
; GFX11-NEXT: s_addc_u32 s8, s8, 0
|
|
; GFX11-NEXT: s_add_u32 s11, s11, 3
|
|
; GFX11-NEXT: s_addc_u32 s10, s10, 0
|
|
; GFX11-NEXT: s_add_u32 s13, s13, 3
|
|
; GFX11-NEXT: s_addc_u32 s12, s12, 0
|
|
; GFX11-NEXT: s_add_u32 s28, s28, 3
|
|
; GFX11-NEXT: s_addc_u32 s29, s29, 0
|
|
; GFX11-NEXT: s_add_u32 s26, s26, 3
|
|
; GFX11-NEXT: s_addc_u32 s27, s27, 0
|
|
; GFX11-NEXT: s_add_u32 s24, s24, 3
|
|
; GFX11-NEXT: s_addc_u32 s25, s25, 0
|
|
; GFX11-NEXT: s_add_u32 s22, s22, 3
|
|
; GFX11-NEXT: s_addc_u32 s23, s23, 0
|
|
; GFX11-NEXT: s_add_u32 s20, s20, 3
|
|
; GFX11-NEXT: s_addc_u32 s21, s21, 0
|
|
; GFX11-NEXT: s_add_u32 s18, s18, 3
|
|
; GFX11-NEXT: s_addc_u32 s19, s19, 0
|
|
; GFX11-NEXT: s_add_u32 s16, s16, 3
|
|
; GFX11-NEXT: s_addc_u32 s17, s17, 0
|
|
; GFX11-NEXT: s_add_u32 s2, s2, 3
|
|
; GFX11-NEXT: s_addc_u32 s3, s3, 0
|
|
; GFX11-NEXT: s_add_u32 s0, s0, 3
|
|
; GFX11-NEXT: s_addc_u32 s1, s1, 0
|
|
; GFX11-NEXT: s_lshr_b32 s14, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s74, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s76, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s77, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s78, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s88, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s0, 16
|
|
; GFX11-NEXT: .LBB45_3: ; %end
|
|
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s88
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s78
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s77
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s76
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s74
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s62
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s60
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s59
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s58
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s57
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s56
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s44
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s40
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s14
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s6
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s5 :: v_dual_mov_b32 v27, s4
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB45_4:
|
|
; GFX11-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-NEXT: ; implicit-def: $sgpr14
|
|
; GFX11-NEXT: s_branch .LBB45_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <14 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <14 x i64> %a1 to <56 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x i64> %a to <56 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x half> %phi
|
|
}
|
|
|
|
define <14 x i64> @bitcast_v56f16_to_v14i64(<56 x half> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v56f16_to_v14i64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_mov_b32_e32 v59, v0
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v22
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v21
|
|
; SI-NEXT: v_mov_b32_e32 v48, v19
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v20
|
|
; SI-NEXT: v_mov_b32_e32 v49, v18
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; SI-NEXT: v_mov_b32_e32 v50, v17
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; SI-NEXT: v_mov_b32_e32 v51, v16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; SI-NEXT: v_mov_b32_e32 v52, v15
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; SI-NEXT: v_mov_b32_e32 v53, v14
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; SI-NEXT: v_mov_b32_e32 v54, v13
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; SI-NEXT: v_mov_b32_e32 v55, v12
|
|
; SI-NEXT: v_mov_b32_e32 v40, v11
|
|
; SI-NEXT: v_mov_b32_e32 v41, v10
|
|
; SI-NEXT: v_mov_b32_e32 v42, v9
|
|
; SI-NEXT: v_mov_b32_e32 v43, v8
|
|
; SI-NEXT: v_mov_b32_e32 v44, v7
|
|
; SI-NEXT: v_mov_b32_e32 v45, v6
|
|
; SI-NEXT: v_mov_b32_e32 v46, v5
|
|
; SI-NEXT: v_mov_b32_e32 v47, v4
|
|
; SI-NEXT: v_mov_b32_e32 v56, v3
|
|
; SI-NEXT: v_mov_b32_e32 v57, v2
|
|
; SI-NEXT: v_mov_b32_e32 v58, v1
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v40
|
|
; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v41
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v42
|
|
; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v43
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v44
|
|
; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v46
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v47
|
|
; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v56
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v57
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v58
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v59
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB46_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v59
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v39
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v58
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v3
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v57
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v62
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v47
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v46
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v45
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v61
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v43
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v60
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v41
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v63
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v11
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v40
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v55
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v53
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v52
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v51
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v50
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v49
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v27
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: .LBB46_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB46_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v32
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v58
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v56
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v47
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v62
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v46
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v45
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v37
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v44
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v43
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v36
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v42
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v41
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v61
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v40
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v55
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; SI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v35
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v54
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v53
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v60
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v52
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v51
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v34
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v49
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v48
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v63
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v33
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v29, 0x38000000, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v16
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v18, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v19
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v21, v19
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v27, v25
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_or_b32_e32 v24, v26, v24
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v27, v29, v27
|
|
; SI-NEXT: .LBB46_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56f16_to_v14i64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_mov_b32_e32 v32, v27
|
|
; VI-NEXT: v_mov_b32_e32 v33, v26
|
|
; VI-NEXT: v_mov_b32_e32 v34, v25
|
|
; VI-NEXT: v_mov_b32_e32 v35, v24
|
|
; VI-NEXT: v_mov_b32_e32 v36, v23
|
|
; VI-NEXT: v_mov_b32_e32 v37, v22
|
|
; VI-NEXT: v_mov_b32_e32 v38, v21
|
|
; VI-NEXT: v_mov_b32_e32 v39, v20
|
|
; VI-NEXT: v_mov_b32_e32 v48, v19
|
|
; VI-NEXT: v_mov_b32_e32 v49, v18
|
|
; VI-NEXT: v_mov_b32_e32 v50, v17
|
|
; VI-NEXT: v_mov_b32_e32 v51, v16
|
|
; VI-NEXT: v_mov_b32_e32 v52, v15
|
|
; VI-NEXT: v_mov_b32_e32 v53, v14
|
|
; VI-NEXT: v_mov_b32_e32 v54, v13
|
|
; VI-NEXT: v_mov_b32_e32 v55, v12
|
|
; VI-NEXT: v_mov_b32_e32 v40, v11
|
|
; VI-NEXT: v_mov_b32_e32 v41, v10
|
|
; VI-NEXT: v_mov_b32_e32 v42, v9
|
|
; VI-NEXT: v_mov_b32_e32 v43, v8
|
|
; VI-NEXT: v_mov_b32_e32 v44, v7
|
|
; VI-NEXT: v_mov_b32_e32 v45, v6
|
|
; VI-NEXT: v_mov_b32_e32 v46, v5
|
|
; VI-NEXT: v_mov_b32_e32 v47, v4
|
|
; VI-NEXT: v_mov_b32_e32 v56, v3
|
|
; VI-NEXT: v_mov_b32_e32 v57, v2
|
|
; VI-NEXT: v_mov_b32_e32 v58, v1
|
|
; VI-NEXT: v_mov_b32_e32 v59, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB46_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_mov_b32_e32 v27, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v27, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v27, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v27, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v27, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v27, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v27, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v27, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v27, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v27, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v27, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v27, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v27, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v27, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v27, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v27, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v27, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v27, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v27, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v27, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v27, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v27, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v24, v27, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v25, v27, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v26, v27, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v27, v27, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v56, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v47, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v46, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v45, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v44, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v43, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v42, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v41, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v40, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v53, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v52, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v51, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v50, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v49, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v39, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v37, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v36, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v35, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v34, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v33, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v32, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: ; implicit-def: $vgpr59
|
|
; VI-NEXT: ; implicit-def: $vgpr58
|
|
; VI-NEXT: ; implicit-def: $vgpr57
|
|
; VI-NEXT: ; implicit-def: $vgpr56
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: .LBB46_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB46_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; VI-NEXT: v_add_f16_sdwa v0, v59, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v59
|
|
; VI-NEXT: v_add_f16_sdwa v2, v58, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v58
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v2, v57, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v57
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v3, v56, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v56
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_add_f16_sdwa v4, v47, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v47
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_add_f16_sdwa v5, v46, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v46
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_add_f16_sdwa v6, v45, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v45
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_add_f16_sdwa v7, v44, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v44
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_add_f16_sdwa v8, v43, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v43
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_add_f16_sdwa v9, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, 0x200, v42
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v41, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, 0x200, v41
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_add_f16_sdwa v11, v40, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, 0x200, v40
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_add_f16_sdwa v12, v55, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, 0x200, v55
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_add_f16_sdwa v13, v54, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v54
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_add_f16_sdwa v14, v53, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v53
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v52, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v52
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v51, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v51
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v50, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v50
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v49, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v48, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v39, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v38, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v37, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v23, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v36, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: v_add_f16_sdwa v24, v35, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v25, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; VI-NEXT: v_add_f16_sdwa v25, v34, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v26, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; VI-NEXT: v_add_f16_sdwa v26, v33, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v26, v28, v26
|
|
; VI-NEXT: v_add_f16_sdwa v27, v32, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; VI-NEXT: .LBB46_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56f16_to_v14i64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_mov_b32_e32 v59, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v22
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v13
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v12
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v11
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v10
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v40
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v56, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v57, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v58, v1
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v41
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v47
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v56
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v57
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v58
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v59
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB46_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v37
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: .LBB46_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB46_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: s_movk_i32 s7, 0x200
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v0, v0, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, v1, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, v2, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, v3, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, v4, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, v5, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, v12, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, v13, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, v14, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, v15, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, v16, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, v17, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, v18, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, v19, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, v20, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, v21, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, v22, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, v23, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v24, v24, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v25, v25, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v26, v26, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v27, v27, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB46_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v14i64:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB46_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: .LBB46_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56f16_to_v14i64:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v69, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v71, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v80, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v68, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v67, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v66, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v65, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v64, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v54, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v53, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v52, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v51, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v50, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v49, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v48, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v39, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v38, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v37, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v36, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v35, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v34, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v33, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v32, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v31, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v30, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v29, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB46_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: .LBB46_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <56 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <56 x half> %a1 to <14 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x half> %a to <14 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x i64> %phi
|
|
}
|
|
|
|
define inreg <14 x i64> @bitcast_v56f16_to_v14i64_scalar(<56 x half> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v56f16_to_v14i64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v32, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v32, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v32, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v32, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v32, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v32, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v32, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v32, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v32, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v32, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v32, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v32, s51, 11
|
|
; SI-NEXT: v_writelane_b32 v32, s52, 12
|
|
; SI-NEXT: v_writelane_b32 v32, s53, 13
|
|
; SI-NEXT: v_writelane_b32 v32, s54, 14
|
|
; SI-NEXT: v_writelane_b32 v32, s55, 15
|
|
; SI-NEXT: v_writelane_b32 v32, s64, 16
|
|
; SI-NEXT: v_writelane_b32 v32, s65, 17
|
|
; SI-NEXT: v_writelane_b32 v32, s66, 18
|
|
; SI-NEXT: v_writelane_b32 v32, s67, 19
|
|
; SI-NEXT: v_writelane_b32 v32, s68, 20
|
|
; SI-NEXT: v_writelane_b32 v32, s69, 21
|
|
; SI-NEXT: v_writelane_b32 v32, s70, 22
|
|
; SI-NEXT: v_writelane_b32 v32, s71, 23
|
|
; SI-NEXT: v_writelane_b32 v32, s80, 24
|
|
; SI-NEXT: v_writelane_b32 v32, s81, 25
|
|
; SI-NEXT: v_writelane_b32 v32, s82, 26
|
|
; SI-NEXT: v_writelane_b32 v32, s83, 27
|
|
; SI-NEXT: v_writelane_b32 v32, s84, 28
|
|
; SI-NEXT: v_writelane_b32 v32, s85, 29
|
|
; SI-NEXT: v_writelane_b32 v32, s86, 30
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s72, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s74, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s76, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s78, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s88, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s90, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s92, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s95, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s34, v0
|
|
; SI-NEXT: v_writelane_b32 v32, s87, 31
|
|
; SI-NEXT: s_lshr_b32 s94, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s69, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s70, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s71, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s80, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s81, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s82, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s83, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s84, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s85, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s86, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s87, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s7, s6, 16
|
|
; SI-NEXT: s_lshr_b32 s9, s8, 16
|
|
; SI-NEXT: s_lshr_b32 s11, s10, 16
|
|
; SI-NEXT: s_lshr_b32 s13, s12, 16
|
|
; SI-NEXT: s_lshr_b32 s15, s14, 16
|
|
; SI-NEXT: s_lshr_b32 s73, s72, 16
|
|
; SI-NEXT: s_lshr_b32 s75, s74, 16
|
|
; SI-NEXT: s_lshr_b32 s77, s76, 16
|
|
; SI-NEXT: s_lshr_b32 s79, s78, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s88, 16
|
|
; SI-NEXT: s_lshr_b32 s91, s90, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s92, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s95, 16
|
|
; SI-NEXT: s_lshr_b32 s68, s34, 16
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB47_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s36, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s37, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s38, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s39, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s40, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s41, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s42, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; SI-NEXT: s_or_b32 s43, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s44, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s45, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s69, 16
|
|
; SI-NEXT: s_or_b32 s46, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s47, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s48, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s94, 16
|
|
; SI-NEXT: s_or_b32 s49, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s34, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s50, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s95, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; SI-NEXT: s_or_b32 s51, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s92, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s52, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s90, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s91, 16
|
|
; SI-NEXT: s_or_b32 s53, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s54, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s78, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s55, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s76, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s77, 16
|
|
; SI-NEXT: s_or_b32 s56, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s74, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; SI-NEXT: s_or_b32 s57, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s72, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s73, 16
|
|
; SI-NEXT: s_or_b32 s58, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; SI-NEXT: s_or_b32 s59, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; SI-NEXT: s_or_b32 s60, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; SI-NEXT: s_or_b32 s61, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s9, 16
|
|
; SI-NEXT: s_or_b32 s62, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; SI-NEXT: s_or_b32 s63, s4, s5
|
|
; SI-NEXT: s_cbranch_execnz .LBB47_4
|
|
; SI-NEXT: .LBB47_2: ; %cmp.true
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, s87
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s86
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s17
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s85
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s18
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s84
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s19
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s83
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s20
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_or_b32_e32 v3, v5, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s82
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_or_b32_e32 v4, v6, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s81
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s22
|
|
; SI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s80
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s23
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_or_b32_e32 v6, v8, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s71
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v9, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s70
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s25
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s69
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s26
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_or_b32_e32 v9, v11, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v12, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s30
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s28
|
|
; SI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s94
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s29
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s68
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v13, v15, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s31
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s95
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s93
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s92
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s91
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v18, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s90
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s89
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s88
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s79
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s78
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s77
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v21, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s76
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s75
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s74
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s73
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s72
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s13
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s12
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s10
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: v_or_b32_e32 v24, v26, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v27, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, s7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, s6
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v29, 0x38000000, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v27, v29, v27
|
|
; SI-NEXT: s_branch .LBB47_5
|
|
; SI-NEXT: .LBB47_3:
|
|
; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; SI-NEXT: s_branch .LBB47_2
|
|
; SI-NEXT: .LBB47_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_mov_b32_e32 v28, s64
|
|
; SI-NEXT: v_mov_b32_e32 v29, s65
|
|
; SI-NEXT: v_mov_b32_e32 v30, s66
|
|
; SI-NEXT: v_mov_b32_e32 v31, s67
|
|
; SI-NEXT: .LBB47_5: ; %end
|
|
; SI-NEXT: v_readlane_b32 s87, v32, 31
|
|
; SI-NEXT: v_readlane_b32 s86, v32, 30
|
|
; SI-NEXT: v_readlane_b32 s85, v32, 29
|
|
; SI-NEXT: v_readlane_b32 s84, v32, 28
|
|
; SI-NEXT: v_readlane_b32 s83, v32, 27
|
|
; SI-NEXT: v_readlane_b32 s82, v32, 26
|
|
; SI-NEXT: v_readlane_b32 s81, v32, 25
|
|
; SI-NEXT: v_readlane_b32 s80, v32, 24
|
|
; SI-NEXT: v_readlane_b32 s71, v32, 23
|
|
; SI-NEXT: v_readlane_b32 s70, v32, 22
|
|
; SI-NEXT: v_readlane_b32 s69, v32, 21
|
|
; SI-NEXT: v_readlane_b32 s68, v32, 20
|
|
; SI-NEXT: v_readlane_b32 s67, v32, 19
|
|
; SI-NEXT: v_readlane_b32 s66, v32, 18
|
|
; SI-NEXT: v_readlane_b32 s65, v32, 17
|
|
; SI-NEXT: v_readlane_b32 s64, v32, 16
|
|
; SI-NEXT: v_readlane_b32 s55, v32, 15
|
|
; SI-NEXT: v_readlane_b32 s54, v32, 14
|
|
; SI-NEXT: v_readlane_b32 s53, v32, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v32, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v32, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v32, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v32, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v32, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v32, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v32, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v32, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v32, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v32, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v32, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v32, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v32, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56f16_to_v14i64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v32, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s31, 1
|
|
; VI-NEXT: v_writelane_b32 v32, s34, 2
|
|
; VI-NEXT: v_writelane_b32 v32, s35, 3
|
|
; VI-NEXT: v_writelane_b32 v32, s36, 4
|
|
; VI-NEXT: v_writelane_b32 v32, s37, 5
|
|
; VI-NEXT: v_writelane_b32 v32, s38, 6
|
|
; VI-NEXT: v_writelane_b32 v32, s39, 7
|
|
; VI-NEXT: v_writelane_b32 v32, s48, 8
|
|
; VI-NEXT: v_writelane_b32 v32, s49, 9
|
|
; VI-NEXT: v_writelane_b32 v32, s50, 10
|
|
; VI-NEXT: v_writelane_b32 v32, s51, 11
|
|
; VI-NEXT: v_writelane_b32 v32, s52, 12
|
|
; VI-NEXT: v_writelane_b32 v32, s53, 13
|
|
; VI-NEXT: v_writelane_b32 v32, s54, 14
|
|
; VI-NEXT: v_writelane_b32 v32, s55, 15
|
|
; VI-NEXT: v_writelane_b32 v32, s64, 16
|
|
; VI-NEXT: v_writelane_b32 v32, s65, 17
|
|
; VI-NEXT: v_writelane_b32 v32, s66, 18
|
|
; VI-NEXT: v_writelane_b32 v32, s67, 19
|
|
; VI-NEXT: v_writelane_b32 v32, s68, 20
|
|
; VI-NEXT: v_writelane_b32 v32, s69, 21
|
|
; VI-NEXT: v_writelane_b32 v32, s70, 22
|
|
; VI-NEXT: v_writelane_b32 v32, s71, 23
|
|
; VI-NEXT: v_writelane_b32 v32, s80, 24
|
|
; VI-NEXT: v_writelane_b32 v32, s81, 25
|
|
; VI-NEXT: v_writelane_b32 v32, s82, 26
|
|
; VI-NEXT: v_writelane_b32 v32, s83, 27
|
|
; VI-NEXT: v_writelane_b32 v32, s84, 28
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; VI-NEXT: v_writelane_b32 v32, s85, 29
|
|
; VI-NEXT: s_lshr_b32 s15, s8, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v11
|
|
; VI-NEXT: ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
|
|
; VI-NEXT: v_writelane_b32 v32, s86, 30
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: s_lshr_b32 s61, s10, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s72, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s74, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s76, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s79, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s34, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s69, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s80, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s83, v0
|
|
; VI-NEXT: v_writelane_b32 v33, s15, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s87, 31
|
|
; VI-NEXT: s_lshr_b32 s56, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s68, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s70, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s81, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s84, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s86, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s87, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s7, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s9, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s11, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s13, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s72, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s74, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s76, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s79, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s91, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s34, 16
|
|
; VI-NEXT: s_lshr_b32 s71, s69, 16
|
|
; VI-NEXT: s_lshr_b32 s82, s80, 16
|
|
; VI-NEXT: s_lshr_b32 s85, s83, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v33, s61, 1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_writelane_b32 v33, s60, 2
|
|
; VI-NEXT: v_writelane_b32 v33, s59, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB47_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_or_b32 s36, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s5, s9, 16
|
|
; VI-NEXT: s_or_b32 s38, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s11, 16
|
|
; VI-NEXT: s_or_b32 s39, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; VI-NEXT: s_or_b32 s37, s40, s41
|
|
; VI-NEXT: s_or_b32 s40, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; VI-NEXT: s_or_b32 s41, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; VI-NEXT: s_or_b32 s42, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; VI-NEXT: s_or_b32 s43, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s44, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s45, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s46, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s47, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_or_b32 s48, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s5, s56, 16
|
|
; VI-NEXT: s_or_b32 s49, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s83
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s50, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s80
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s51, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s69
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s52, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s34
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s53, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s91
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s54, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s79
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s55, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s76
|
|
; VI-NEXT: s_lshl_b32 s5, s62, 16
|
|
; VI-NEXT: s_mov_b32 s77, s56
|
|
; VI-NEXT: s_or_b32 s56, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s74
|
|
; VI-NEXT: s_lshl_b32 s5, s57, 16
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: s_mov_b32 s75, s57
|
|
; VI-NEXT: s_or_b32 s57, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s72
|
|
; VI-NEXT: s_lshl_b32 s5, s58, 16
|
|
; VI-NEXT: s_mov_b32 s78, s58
|
|
; VI-NEXT: s_or_b32 s58, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s5, s59, 16
|
|
; VI-NEXT: s_or_b32 s59, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s5, s60, 16
|
|
; VI-NEXT: s_or_b32 s60, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s5, s61, 16
|
|
; VI-NEXT: s_or_b32 s61, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_mov_b32 s88, s62
|
|
; VI-NEXT: s_or_b32 s62, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s5, s63, 16
|
|
; VI-NEXT: s_mov_b32 s15, s63
|
|
; VI-NEXT: s_or_b32 s63, s4, s5
|
|
; VI-NEXT: s_cbranch_execnz .LBB47_4
|
|
; VI-NEXT: .LBB47_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; VI-NEXT: v_mov_b32_e32 v0, s13
|
|
; VI-NEXT: v_mov_b32_e32 v2, s11
|
|
; VI-NEXT: v_add_f16_sdwa v0, v0, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, s16, v27
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s17, v27
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v2, s9
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s18, v27
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v3, s7
|
|
; VI-NEXT: v_add_f16_sdwa v3, v3, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, s19, v27
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_mov_b32_e32 v4, s87
|
|
; VI-NEXT: v_add_f16_sdwa v4, v4, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, s20, v27
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_mov_b32_e32 v5, s86
|
|
; VI-NEXT: v_add_f16_sdwa v5, v5, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, s21, v27
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_mov_b32_e32 v6, s84
|
|
; VI-NEXT: v_add_f16_sdwa v6, v6, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, s22, v27
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_mov_b32_e32 v7, s81
|
|
; VI-NEXT: v_add_f16_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, s23, v27
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_mov_b32_e32 v8, s70
|
|
; VI-NEXT: v_add_f16_sdwa v8, v8, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, s24, v27
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_mov_b32_e32 v9, s68
|
|
; VI-NEXT: v_add_f16_sdwa v9, v9, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, s25, v27
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_mov_b32_e32 v10, s31
|
|
; VI-NEXT: v_add_f16_sdwa v10, v10, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, s26, v27
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_mov_b32_e32 v11, s90
|
|
; VI-NEXT: v_add_f16_sdwa v11, v11, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, s27, v27
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_mov_b32_e32 v12, s73
|
|
; VI-NEXT: v_add_f16_sdwa v12, v12, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, s28, v27
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_mov_b32_e32 v13, s77
|
|
; VI-NEXT: v_add_f16_sdwa v13, v13, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, s29, v27
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_mov_b32_e32 v14, s85
|
|
; VI-NEXT: v_add_f16_sdwa v14, v14, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, s83, v27
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_mov_b32_e32 v15, s82
|
|
; VI-NEXT: v_add_f16_sdwa v15, v15, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, s80, v27
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_mov_b32_e32 v16, s71
|
|
; VI-NEXT: v_add_f16_sdwa v16, v16, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, s69, v27
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_mov_b32_e32 v17, s35
|
|
; VI-NEXT: v_add_f16_sdwa v17, v17, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, s34, v27
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_mov_b32_e32 v18, s30
|
|
; VI-NEXT: v_add_f16_sdwa v18, v18, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, s91, v27
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_mov_b32_e32 v19, s89
|
|
; VI-NEXT: v_add_f16_sdwa v19, v19, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, s79, v27
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_mov_b32_e32 v20, s88
|
|
; VI-NEXT: v_add_f16_sdwa v20, v20, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, s76, v27
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_mov_b32_e32 v21, s75
|
|
; VI-NEXT: v_add_f16_sdwa v21, v21, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, s74, v27
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_mov_b32_e32 v22, s78
|
|
; VI-NEXT: v_add_f16_sdwa v22, v22, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v23, s72, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 3
|
|
; VI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; VI-NEXT: v_mov_b32_e32 v23, s4
|
|
; VI-NEXT: v_add_f16_sdwa v23, v23, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, s14, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 2
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: v_mov_b32_e32 v24, s4
|
|
; VI-NEXT: v_add_f16_sdwa v24, v24, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v25, s12, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 1
|
|
; VI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; VI-NEXT: v_mov_b32_e32 v25, s4
|
|
; VI-NEXT: v_add_f16_sdwa v25, v25, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v26, s10, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 0
|
|
; VI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; VI-NEXT: v_mov_b32_e32 v26, s4
|
|
; VI-NEXT: v_add_f16_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, s8, v27
|
|
; VI-NEXT: v_or_b32_e32 v26, v28, v26
|
|
; VI-NEXT: v_mov_b32_e32 v28, s15
|
|
; VI-NEXT: v_add_f16_sdwa v28, v28, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v27, s6, v27
|
|
; VI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; VI-NEXT: s_branch .LBB47_5
|
|
; VI-NEXT: .LBB47_3:
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: s_mov_b32 s88, s62
|
|
; VI-NEXT: s_mov_b32 s77, s56
|
|
; VI-NEXT: s_mov_b32 s75, s57
|
|
; VI-NEXT: s_mov_b32 s78, s58
|
|
; VI-NEXT: s_mov_b32 s15, s63
|
|
; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; VI-NEXT: s_branch .LBB47_2
|
|
; VI-NEXT: .LBB47_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_mov_b32_e32 v28, s64
|
|
; VI-NEXT: v_mov_b32_e32 v29, s65
|
|
; VI-NEXT: v_mov_b32_e32 v30, s66
|
|
; VI-NEXT: v_mov_b32_e32 v31, s67
|
|
; VI-NEXT: .LBB47_5: ; %end
|
|
; VI-NEXT: v_readlane_b32 s87, v32, 31
|
|
; VI-NEXT: v_readlane_b32 s86, v32, 30
|
|
; VI-NEXT: v_readlane_b32 s85, v32, 29
|
|
; VI-NEXT: v_readlane_b32 s84, v32, 28
|
|
; VI-NEXT: v_readlane_b32 s83, v32, 27
|
|
; VI-NEXT: v_readlane_b32 s82, v32, 26
|
|
; VI-NEXT: v_readlane_b32 s81, v32, 25
|
|
; VI-NEXT: v_readlane_b32 s80, v32, 24
|
|
; VI-NEXT: v_readlane_b32 s71, v32, 23
|
|
; VI-NEXT: v_readlane_b32 s70, v32, 22
|
|
; VI-NEXT: v_readlane_b32 s69, v32, 21
|
|
; VI-NEXT: v_readlane_b32 s68, v32, 20
|
|
; VI-NEXT: v_readlane_b32 s67, v32, 19
|
|
; VI-NEXT: v_readlane_b32 s66, v32, 18
|
|
; VI-NEXT: v_readlane_b32 s65, v32, 17
|
|
; VI-NEXT: v_readlane_b32 s64, v32, 16
|
|
; VI-NEXT: v_readlane_b32 s55, v32, 15
|
|
; VI-NEXT: v_readlane_b32 s54, v32, 14
|
|
; VI-NEXT: v_readlane_b32 s53, v32, 13
|
|
; VI-NEXT: v_readlane_b32 s52, v32, 12
|
|
; VI-NEXT: v_readlane_b32 s51, v32, 11
|
|
; VI-NEXT: v_readlane_b32 s50, v32, 10
|
|
; VI-NEXT: v_readlane_b32 s49, v32, 9
|
|
; VI-NEXT: v_readlane_b32 s48, v32, 8
|
|
; VI-NEXT: v_readlane_b32 s39, v32, 7
|
|
; VI-NEXT: v_readlane_b32 s38, v32, 6
|
|
; VI-NEXT: v_readlane_b32 s37, v32, 5
|
|
; VI-NEXT: v_readlane_b32 s36, v32, 4
|
|
; VI-NEXT: v_readlane_b32 s35, v32, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v32, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v32, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v32, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56f16_to_v14i64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: s_lshr_b32 s12, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s13, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s14, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s15, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s88, v5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v3
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v2
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: s_lshr_b32 s4, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s5, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s6, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s7, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s8, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s9, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s10, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s63, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s62, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s61, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s60, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s59, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s58, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s57, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s56, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s88, 16
|
|
; GFX9-NEXT: s_lshr_b32 s17, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s18, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s19, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s20, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s21, s12, 16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_cmp_lg_u32 s42, 0
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s50, s12, s21
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s51, s13, s20
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s52, s14, s19
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s53, s15, s18
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s54, s16, s17
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s55, s88, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s56, s56, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s57, s57, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s58, s58, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s59, s59, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s60, s60, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s61, s61, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s62, s62, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s63, s63, s72
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB47_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB47_4
|
|
; GFX9-NEXT: .LBB47_2: ; %cmp.true
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; GFX9-NEXT: v_pk_add_f16 v0, s36, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, s37, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, s38, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, s39, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, s40, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, s41, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, s42, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, s43, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, s44, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v9, s45, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, s46, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, s47, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, s48, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, s49, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, s50, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, s51, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, s52, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, s53, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, s54, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, s55, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, s56, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, s57, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, s58, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, s59, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v24, s60, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v25, s61, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v26, s62, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v27, s63, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_branch .LBB47_5
|
|
; GFX9-NEXT: .LBB47_3:
|
|
; GFX9-NEXT: s_branch .LBB47_2
|
|
; GFX9-NEXT: .LBB47_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB47_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v56f16_to_v14i64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s44, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s74, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s76, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s77, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s78, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s88, v10
|
|
; GFX11-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s0, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s44, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s56, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s58, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s59, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s60, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s62, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s74, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s76, 16
|
|
; GFX11-NEXT: s_lshr_b32 s90, s77, 16
|
|
; GFX11-NEXT: s_lshr_b32 s91, s78, 16
|
|
; GFX11-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s88, 0
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s78, s91
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s77, s90
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s76, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s74, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s62, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s60, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s59, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s58, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s56, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s44, s57
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB47_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB47_4
|
|
; GFX11-NEXT: .LBB47_2: ; %cmp.true
|
|
; GFX11-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v18, 0x200, s18 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v19, 0x200, s19 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v20, 0x200, s20 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v21, 0x200, s21 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v22, 0x200, s22 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v23, 0x200, s23 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v24, 0x200, s24 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v25, 0x200, s25 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v26, 0x200, s26 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v27, 0x200, s27 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB47_3:
|
|
; GFX11-NEXT: s_branch .LBB47_2
|
|
; GFX11-NEXT: .LBB47_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <56 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <56 x half> %a1 to <14 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x half> %a to <14 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x i64> %phi
|
|
}
|
|
|
|
define <56 x i16> @bitcast_v14f64_to_v56i16(<14 x double> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v14f64_to_v56i16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB48_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v37, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v39, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB48_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB48_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; SI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; SI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; SI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; SI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; SI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; SI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; SI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; SI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; SI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; SI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; SI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; SI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; SI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v37, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v39, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB48_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v47
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v40
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v41
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v53
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v50
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v39
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v37
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14f64_to_v56i16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB48_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB48_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB48_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB48_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14f64_to_v56i16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB48_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB48_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB48_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB48_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v47, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v46, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v45, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v44, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v43, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v42, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v41, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v40, v7, s4
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v8, v55, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v54, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v53, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v52, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v51, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v49, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v48, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v39, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v38, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v37, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v36, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v35, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v34, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v33, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v32, v23, s4
|
|
; GFX9-NEXT: v_perm_b32 v24, v31, v24, s4
|
|
; GFX9-NEXT: v_perm_b32 v25, v30, v25, s4
|
|
; GFX9-NEXT: v_perm_b32 v26, v29, v26, s4
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s4
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v14f64_to_v56i16:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB48_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX11-TRUE16-NEXT: .LBB48_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v14f64_to_v56i16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr67
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB48_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB48_2: ; %Flow
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB48_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB48_4: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v71, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v69, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v68, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v67, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v66, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v65, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v64, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v55, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v54, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v53, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v52, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v51, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v49, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v39, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v37, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v36, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v35, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v34, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v33, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v32, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v31, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v30, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v29, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <14 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <14 x double> %a1 to <56 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x double> %a to <56 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x i16> %phi
|
|
}
|
|
|
|
define inreg <56 x i16> @bitcast_v14f64_to_v56i16_scalar(<14 x double> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v14f64_to_v56i16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v58, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v58, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v58, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v58, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v58, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v58, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v58, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v58, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v58, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v58, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v58, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v58, s51, 11
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_writelane_b32 v58, s52, 12
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s5, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v0
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_writelane_b32 v58, s53, 13
|
|
; SI-NEXT: s_cbranch_scc0 .LBB49_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s53, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB49_4
|
|
; SI-NEXT: .LBB49_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[26:27], s[40:41], 1.0
|
|
; SI-NEXT: v_add_f64 v[24:25], s[14:15], 1.0
|
|
; SI-NEXT: v_add_f64 v[22:23], s[12:13], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[28:29], v[26:27], 16
|
|
; SI-NEXT: v_add_f64 v[20:21], s[10:11], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[29:30], v[24:25], 16
|
|
; SI-NEXT: v_add_f64 v[18:19], s[8:9], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[30:31], v[22:23], 16
|
|
; SI-NEXT: v_add_f64 v[16:17], s[6:7], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[31:32], v[20:21], 16
|
|
; SI-NEXT: v_add_f64 v[14:15], s[4:5], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[32:33], v[18:19], 16
|
|
; SI-NEXT: v_add_f64 v[12:13], s[28:29], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[33:34], v[16:17], 16
|
|
; SI-NEXT: v_add_f64 v[10:11], s[26:27], 1.0
|
|
; SI-NEXT: v_add_f64 v[4:5], s[20:21], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[34:35], v[14:15], 16
|
|
; SI-NEXT: v_add_f64 v[8:9], s[24:25], 1.0
|
|
; SI-NEXT: v_add_f64 v[2:3], s[18:19], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[35:36], v[12:13], 16
|
|
; SI-NEXT: v_add_f64 v[6:7], s[22:23], 1.0
|
|
; SI-NEXT: v_add_f64 v[0:1], s[16:17], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[36:37], v[10:11], 16
|
|
; SI-NEXT: v_lshr_b64 v[48:49], v[4:5], 16
|
|
; SI-NEXT: v_lshr_b64 v[37:38], v[8:9], 16
|
|
; SI-NEXT: v_lshr_b64 v[49:50], v[2:3], 16
|
|
; SI-NEXT: v_lshr_b64 v[38:39], v[6:7], 16
|
|
; SI-NEXT: v_lshr_b64 v[50:51], v[0:1], 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v1
|
|
; SI-NEXT: s_branch .LBB49_5
|
|
; SI-NEXT: .LBB49_3:
|
|
; SI-NEXT: ; implicit-def: $sgpr92
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: ; implicit-def: $sgpr31
|
|
; SI-NEXT: ; implicit-def: $sgpr88
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr78
|
|
; SI-NEXT: ; implicit-def: $sgpr35
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr36
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr37
|
|
; SI-NEXT: ; implicit-def: $sgpr72
|
|
; SI-NEXT: ; implicit-def: $sgpr38
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr39
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr48
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr49
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr50
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr51
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr52
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr53
|
|
; SI-NEXT: s_branch .LBB49_2
|
|
; SI-NEXT: .LBB49_4:
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v27, s41
|
|
; SI-NEXT: v_mov_b32_e32 v25, s15
|
|
; SI-NEXT: v_mov_b32_e32 v23, s13
|
|
; SI-NEXT: v_mov_b32_e32 v21, s11
|
|
; SI-NEXT: v_mov_b32_e32 v19, s9
|
|
; SI-NEXT: v_mov_b32_e32 v17, s7
|
|
; SI-NEXT: v_mov_b32_e32 v15, s5
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v14, s4
|
|
; SI-NEXT: v_mov_b32_e32 v16, s6
|
|
; SI-NEXT: v_mov_b32_e32 v18, s8
|
|
; SI-NEXT: v_mov_b32_e32 v20, s10
|
|
; SI-NEXT: v_mov_b32_e32 v22, s12
|
|
; SI-NEXT: v_mov_b32_e32 v24, s14
|
|
; SI-NEXT: v_mov_b32_e32 v26, s40
|
|
; SI-NEXT: v_mov_b32_e32 v52, s53
|
|
; SI-NEXT: v_mov_b32_e32 v53, s52
|
|
; SI-NEXT: v_mov_b32_e32 v54, s51
|
|
; SI-NEXT: v_mov_b32_e32 v55, s50
|
|
; SI-NEXT: v_mov_b32_e32 v40, s49
|
|
; SI-NEXT: v_mov_b32_e32 v41, s48
|
|
; SI-NEXT: v_mov_b32_e32 v42, s39
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_mov_b32_e32 v43, s38
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_mov_b32_e32 v44, s37
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_mov_b32_e32 v45, s36
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_mov_b32_e32 v46, s35
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v47, s34
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_mov_b32_e32 v56, s31
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v57, s30
|
|
; SI-NEXT: v_mov_b32_e32 v50, s92
|
|
; SI-NEXT: v_mov_b32_e32 v49, s90
|
|
; SI-NEXT: v_mov_b32_e32 v48, s88
|
|
; SI-NEXT: v_mov_b32_e32 v38, s78
|
|
; SI-NEXT: v_mov_b32_e32 v37, s76
|
|
; SI-NEXT: v_mov_b32_e32 v36, s74
|
|
; SI-NEXT: v_mov_b32_e32 v35, s72
|
|
; SI-NEXT: v_mov_b32_e32 v34, s62
|
|
; SI-NEXT: v_mov_b32_e32 v33, s60
|
|
; SI-NEXT: v_mov_b32_e32 v32, s58
|
|
; SI-NEXT: v_mov_b32_e32 v31, s56
|
|
; SI-NEXT: v_mov_b32_e32 v30, s46
|
|
; SI-NEXT: v_mov_b32_e32 v29, s44
|
|
; SI-NEXT: v_mov_b32_e32 v28, s42
|
|
; SI-NEXT: .LBB49_5: ; %end
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v50
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v39
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v57
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v49
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v39
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v56
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v48
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v41
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v40
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v39
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v38
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v37
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: v_readlane_b32 s53, v58, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v58, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v58, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v58, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v58, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v58, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v58, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v58, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v58, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v58, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v58, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v58, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v58, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v58, 0
|
|
; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14f64_to_v56i16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v56, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v56, s31, 1
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v56, s34, 2
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s5, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v0
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_writelane_b32 v56, s35, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB49_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s44, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s5, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s4, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB49_4
|
|
; VI-NEXT: .LBB49_2: ; %cmp.true
|
|
; VI-NEXT: v_add_f64 v[26:27], s[8:9], 1.0
|
|
; VI-NEXT: v_add_f64 v[24:25], s[10:11], 1.0
|
|
; VI-NEXT: v_add_f64 v[22:23], s[12:13], 1.0
|
|
; VI-NEXT: v_add_f64 v[20:21], s[14:15], 1.0
|
|
; VI-NEXT: v_add_f64 v[18:19], s[40:41], 1.0
|
|
; VI-NEXT: v_add_f64 v[16:17], s[6:7], 1.0
|
|
; VI-NEXT: v_add_f64 v[14:15], s[4:5], 1.0
|
|
; VI-NEXT: v_add_f64 v[12:13], s[28:29], 1.0
|
|
; VI-NEXT: v_add_f64 v[10:11], s[26:27], 1.0
|
|
; VI-NEXT: v_add_f64 v[8:9], s[24:25], 1.0
|
|
; VI-NEXT: v_add_f64 v[6:7], s[22:23], 1.0
|
|
; VI-NEXT: v_add_f64 v[4:5], s[20:21], 1.0
|
|
; VI-NEXT: v_add_f64 v[2:3], s[18:19], 1.0
|
|
; VI-NEXT: v_add_f64 v[0:1], s[16:17], 1.0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v0
|
|
; VI-NEXT: s_branch .LBB49_5
|
|
; VI-NEXT: .LBB49_3:
|
|
; VI-NEXT: ; implicit-def: $sgpr35
|
|
; VI-NEXT: ; implicit-def: $sgpr73
|
|
; VI-NEXT: ; implicit-def: $sgpr34
|
|
; VI-NEXT: ; implicit-def: $sgpr72
|
|
; VI-NEXT: ; implicit-def: $sgpr31
|
|
; VI-NEXT: ; implicit-def: $sgpr63
|
|
; VI-NEXT: ; implicit-def: $sgpr30
|
|
; VI-NEXT: ; implicit-def: $sgpr62
|
|
; VI-NEXT: ; implicit-def: $sgpr91
|
|
; VI-NEXT: ; implicit-def: $sgpr61
|
|
; VI-NEXT: ; implicit-def: $sgpr90
|
|
; VI-NEXT: ; implicit-def: $sgpr60
|
|
; VI-NEXT: ; implicit-def: $sgpr89
|
|
; VI-NEXT: ; implicit-def: $sgpr59
|
|
; VI-NEXT: ; implicit-def: $sgpr88
|
|
; VI-NEXT: ; implicit-def: $sgpr58
|
|
; VI-NEXT: ; implicit-def: $sgpr79
|
|
; VI-NEXT: ; implicit-def: $sgpr57
|
|
; VI-NEXT: ; implicit-def: $sgpr78
|
|
; VI-NEXT: ; implicit-def: $sgpr56
|
|
; VI-NEXT: ; implicit-def: $sgpr77
|
|
; VI-NEXT: ; implicit-def: $sgpr47
|
|
; VI-NEXT: ; implicit-def: $sgpr76
|
|
; VI-NEXT: ; implicit-def: $sgpr46
|
|
; VI-NEXT: ; implicit-def: $sgpr75
|
|
; VI-NEXT: ; implicit-def: $sgpr45
|
|
; VI-NEXT: ; implicit-def: $sgpr74
|
|
; VI-NEXT: ; implicit-def: $sgpr44
|
|
; VI-NEXT: s_branch .LBB49_2
|
|
; VI-NEXT: .LBB49_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v14, s4
|
|
; VI-NEXT: v_mov_b32_e32 v16, s6
|
|
; VI-NEXT: v_mov_b32_e32 v18, s40
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v15, s5
|
|
; VI-NEXT: v_mov_b32_e32 v17, s7
|
|
; VI-NEXT: v_mov_b32_e32 v19, s41
|
|
; VI-NEXT: v_mov_b32_e32 v20, s14
|
|
; VI-NEXT: v_mov_b32_e32 v21, s15
|
|
; VI-NEXT: v_mov_b32_e32 v22, s12
|
|
; VI-NEXT: v_mov_b32_e32 v23, s13
|
|
; VI-NEXT: v_mov_b32_e32 v24, s10
|
|
; VI-NEXT: v_mov_b32_e32 v25, s11
|
|
; VI-NEXT: v_mov_b32_e32 v26, s8
|
|
; VI-NEXT: v_mov_b32_e32 v27, s9
|
|
; VI-NEXT: v_mov_b32_e32 v44, s35
|
|
; VI-NEXT: v_mov_b32_e32 v42, s34
|
|
; VI-NEXT: v_mov_b32_e32 v40, s31
|
|
; VI-NEXT: v_mov_b32_e32 v54, s30
|
|
; VI-NEXT: v_mov_b32_e32 v52, s91
|
|
; VI-NEXT: v_mov_b32_e32 v50, s90
|
|
; VI-NEXT: v_mov_b32_e32 v48, s89
|
|
; VI-NEXT: v_mov_b32_e32 v38, s88
|
|
; VI-NEXT: v_mov_b32_e32 v36, s79
|
|
; VI-NEXT: v_mov_b32_e32 v34, s78
|
|
; VI-NEXT: v_mov_b32_e32 v32, s77
|
|
; VI-NEXT: v_mov_b32_e32 v30, s76
|
|
; VI-NEXT: v_mov_b32_e32 v29, s75
|
|
; VI-NEXT: v_mov_b32_e32 v28, s74
|
|
; VI-NEXT: v_mov_b32_e32 v47, s73
|
|
; VI-NEXT: v_mov_b32_e32 v46, s72
|
|
; VI-NEXT: v_mov_b32_e32 v45, s63
|
|
; VI-NEXT: v_mov_b32_e32 v43, s62
|
|
; VI-NEXT: v_mov_b32_e32 v41, s61
|
|
; VI-NEXT: v_mov_b32_e32 v55, s60
|
|
; VI-NEXT: v_mov_b32_e32 v53, s59
|
|
; VI-NEXT: v_mov_b32_e32 v51, s58
|
|
; VI-NEXT: v_mov_b32_e32 v49, s57
|
|
; VI-NEXT: v_mov_b32_e32 v39, s56
|
|
; VI-NEXT: v_mov_b32_e32 v37, s47
|
|
; VI-NEXT: v_mov_b32_e32 v35, s46
|
|
; VI-NEXT: v_mov_b32_e32 v33, s45
|
|
; VI-NEXT: v_mov_b32_e32 v31, s44
|
|
; VI-NEXT: .LBB49_5: ; %end
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v47
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v46
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v43
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v41
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v55
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v53
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v51
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v49
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v39
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v37
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v35
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v33
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v31
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_readlane_b32 s35, v56, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v56, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v56, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v56, 0
|
|
; VI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14f64_to_v56i16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s5, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v0
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB49_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s44, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s5, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s4, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB49_4
|
|
; GFX9-NEXT: .LBB49_2: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f64 v[26:27], s[8:9], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[24:25], s[10:11], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[22:23], s[12:13], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[20:21], s[14:15], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[18:19], s[40:41], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[16:17], s[6:7], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[14:15], s[4:5], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[12:13], s[28:29], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[10:11], s[26:27], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[8:9], s[24:25], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[6:7], s[22:23], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[4:5], s[20:21], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[2:3], s[18:19], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[0:1], s[16:17], 1.0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v0
|
|
; GFX9-NEXT: s_branch .LBB49_5
|
|
; GFX9-NEXT: .LBB49_3:
|
|
; GFX9-NEXT: ; implicit-def: $sgpr95
|
|
; GFX9-NEXT: ; implicit-def: $sgpr73
|
|
; GFX9-NEXT: ; implicit-def: $sgpr94
|
|
; GFX9-NEXT: ; implicit-def: $sgpr72
|
|
; GFX9-NEXT: ; implicit-def: $sgpr93
|
|
; GFX9-NEXT: ; implicit-def: $sgpr63
|
|
; GFX9-NEXT: ; implicit-def: $sgpr92
|
|
; GFX9-NEXT: ; implicit-def: $sgpr62
|
|
; GFX9-NEXT: ; implicit-def: $sgpr91
|
|
; GFX9-NEXT: ; implicit-def: $sgpr61
|
|
; GFX9-NEXT: ; implicit-def: $sgpr90
|
|
; GFX9-NEXT: ; implicit-def: $sgpr60
|
|
; GFX9-NEXT: ; implicit-def: $sgpr89
|
|
; GFX9-NEXT: ; implicit-def: $sgpr59
|
|
; GFX9-NEXT: ; implicit-def: $sgpr88
|
|
; GFX9-NEXT: ; implicit-def: $sgpr58
|
|
; GFX9-NEXT: ; implicit-def: $sgpr79
|
|
; GFX9-NEXT: ; implicit-def: $sgpr57
|
|
; GFX9-NEXT: ; implicit-def: $sgpr78
|
|
; GFX9-NEXT: ; implicit-def: $sgpr56
|
|
; GFX9-NEXT: ; implicit-def: $sgpr77
|
|
; GFX9-NEXT: ; implicit-def: $sgpr47
|
|
; GFX9-NEXT: ; implicit-def: $sgpr76
|
|
; GFX9-NEXT: ; implicit-def: $sgpr46
|
|
; GFX9-NEXT: ; implicit-def: $sgpr75
|
|
; GFX9-NEXT: ; implicit-def: $sgpr45
|
|
; GFX9-NEXT: ; implicit-def: $sgpr74
|
|
; GFX9-NEXT: ; implicit-def: $sgpr44
|
|
; GFX9-NEXT: s_branch .LBB49_2
|
|
; GFX9-NEXT: .LBB49_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s4
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s5
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, s95
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, s94
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, s93
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, s92
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, s91
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, s90
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, s89
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, s88
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, s79
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, s78
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, s77
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s76
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s75
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s74
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, s73
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, s72
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s44
|
|
; GFX9-NEXT: .LBB49_5: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v0, v44, 16, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v1, v47, 16, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v2, v42, 16, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v3, v46, 16, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v4, v40, 16, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v5, v45, 16, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v7, v43, 16, v7
|
|
; GFX9-NEXT: v_lshl_or_b32 v9, v41, 16, v9
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX9-NEXT: v_lshl_or_b32 v6, v54, 16, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v8, v52, 16, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v50, 16, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v55, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v48, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v53, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v38, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v51, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v36, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v49, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v34, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v39, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v32, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v37, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v30, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v35, 16, v23
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v29, 16, v24
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v33, 16, v25
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v28, 16, v26
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v31, 16, v27
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v14f64_to_v56i16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s11, v9
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s10, v8
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s9, v7
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s8, v6
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s7, v5
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s6, v4
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v3
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v2
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB49_3
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s11, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s61, s10, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s9, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s62, s8, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s63, s6, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s5, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s72, s4, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s13, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s74, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s75, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s76, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s77, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s78, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s57, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s79, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s58, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s88, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s59, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s89, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s60, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s90, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB49_4
|
|
; GFX11-TRUE16-NEXT: .LBB49_2: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[26:27], s[10:11], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[24:25], s[8:9], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[22:23], s[6:7], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[20:21], s[4:5], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[18:19], s[12:13], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[16:17], s[28:29], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[14:15], s[26:27], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[12:13], s[24:25], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[10:11], s[22:23], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[8:9], s[20:21], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[6:7], s[18:19], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[4:5], s[16:17], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[2:3], s[2:3], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[0:1], s[0:1], 1.0
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v27
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v26
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v25
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v24
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v65, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v67, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 16, v0
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB49_5
|
|
; GFX11-TRUE16-NEXT: .LBB49_3:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr90
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB49_2
|
|
; GFX11-TRUE16-NEXT: .LBB49_4:
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, s12 :: v_dual_mov_b32 v19, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s4 :: v_dual_mov_b32 v21, s5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, s6 :: v_dual_mov_b32 v23, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v24, s8 :: v_dual_mov_b32 v25, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, s10 :: v_dual_mov_b32 v27, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v68, s90 :: v_dual_mov_b32 v29, s62
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v66, s89 :: v_dual_mov_b32 v71, s60
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, s88 :: v_dual_mov_b32 v69, s58
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s79 :: v_dual_mov_b32 v67, s57
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v52, s78 :: v_dual_mov_b32 v65, s56
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v50, s77 :: v_dual_mov_b32 v55, s47
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v48, s76 :: v_dual_mov_b32 v53, s46
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v38, s75 :: v_dual_mov_b32 v51, s45
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v36, s74 :: v_dual_mov_b32 v49, s44
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v34, s73 :: v_dual_mov_b32 v39, s43
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v32, s72 :: v_dual_mov_b32 v37, s42
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, s63 :: v_dual_mov_b32 v35, s41
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, s61 :: v_dual_mov_b32 v33, s40
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v70, s59 :: v_dual_mov_b32 v31, s15
|
|
; GFX11-TRUE16-NEXT: .LBB49_5: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v68.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v71.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v66.l
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v70.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v64.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v69.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v67.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v65.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v28.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v31.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v14f64_to_v56i16_scalar:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s11, v9
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s10, v8
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s9, v7
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s8, v6
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s7, v5
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s6, v4
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s5, v3
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s4, v2
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB49_3
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s11, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s61, s10, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s9, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s62, s8, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s63, s6, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s5, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s72, s4, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s13, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s74, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s75, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s76, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s77, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s78, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s57, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s79, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s58, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s88, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s59, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s89, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s60, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s90, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB49_4
|
|
; GFX11-FAKE16-NEXT: .LBB49_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[23:24], s[10:11], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[25:26], s[8:9], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[27:28], s[6:7], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[18:19], s[4:5], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[20:21], s[12:13], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[13:14], s[28:29], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[15:16], s[26:27], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[29:30], s[24:25], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[8:9], s[22:23], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[10:11], s[20:21], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[3:4], s[18:19], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[5:6], s[16:17], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[31:32], s[2:3], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[0:1], s[0:1], 1.0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v28
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v17, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v30
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v12, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v7, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v32
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v2, 16, v31
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB49_5
|
|
; GFX11-FAKE16-NEXT: .LBB49_3:
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr90
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB49_2
|
|
; GFX11-FAKE16-NEXT: .LBB49_4:
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v31, s2
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v10, s20
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v8, s22
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v29, s24 :: v_dual_mov_b32 v20, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v18, s4
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s28 :: v_dual_mov_b32 v32, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v27, s6 :: v_dual_mov_b32 v6, s17
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v25, s8 :: v_dual_mov_b32 v4, s19
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v23, s10 :: v_dual_mov_b32 v30, s25
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v16, s27
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v14, s29
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v28, s7
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v21, s13 :: v_dual_mov_b32 v26, s9
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v19, s5 :: v_dual_mov_b32 v24, s11
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v69, s90 :: v_dual_mov_b32 v2, s89
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v66, s88 :: v_dual_mov_b32 v7, s79
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v55, s78 :: v_dual_mov_b32 v12, s76
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v53, s77 :: v_dual_mov_b32 v50, s75
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s74 :: v_dual_mov_b32 v22, s63
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v39, s73 :: v_dual_mov_b32 v34, s62
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, s72 :: v_dual_mov_b32 v70, s59
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v33, s61 :: v_dual_mov_b32 v68, s58
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v71, s60 :: v_dual_mov_b32 v64, s47
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v67, s57 :: v_dual_mov_b32 v54, s46
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v65, s56 :: v_dual_mov_b32 v52, s45
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v51, s44 :: v_dual_mov_b32 v48, s42
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v49, s43 :: v_dual_mov_b32 v38, s41
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, s40 :: v_dual_mov_b32 v35, s15
|
|
; GFX11-FAKE16-NEXT: .LBB49_5: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v31
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v2, 16, v31
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v32
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v12, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v30
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v70, 16, v31
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xffff, v13
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v54, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v69, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v68, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v7, 16, v32
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v55, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v53, 16, v31
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v50, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v52, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v17, 16, v30
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xffff, v19
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v39, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v37, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v22, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xffff, v28
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xffff, v23
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v71, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v67, 16, v69
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v65, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v64, 16, v32
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v51, 16, v31
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v49, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v48, 16, v30
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v38, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v34, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v36, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v33, 16, v28
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v35, 16, v29
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <14 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <14 x double> %a1 to <56 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x double> %a to <56 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x i16> %phi
|
|
}
|
|
|
|
define <14 x double> @bitcast_v56i16_to_v14f64(<56 x i16> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v56i16_to_v14f64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_mov_b32_e32 v48, v19
|
|
; SI-NEXT: v_mov_b32_e32 v49, v18
|
|
; SI-NEXT: v_mov_b32_e32 v50, v17
|
|
; SI-NEXT: v_mov_b32_e32 v51, v16
|
|
; SI-NEXT: v_mov_b32_e32 v52, v15
|
|
; SI-NEXT: v_mov_b32_e32 v53, v14
|
|
; SI-NEXT: v_mov_b32_e32 v54, v13
|
|
; SI-NEXT: v_mov_b32_e32 v55, v12
|
|
; SI-NEXT: v_mov_b32_e32 v40, v11
|
|
; SI-NEXT: v_mov_b32_e32 v41, v10
|
|
; SI-NEXT: v_mov_b32_e32 v42, v9
|
|
; SI-NEXT: v_mov_b32_e32 v43, v8
|
|
; SI-NEXT: v_mov_b32_e32 v44, v7
|
|
; SI-NEXT: v_mov_b32_e32 v45, v6
|
|
; SI-NEXT: v_mov_b32_e32 v46, v5
|
|
; SI-NEXT: v_mov_b32_e32 v47, v4
|
|
; SI-NEXT: v_mov_b32_e32 v56, v3
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_mov_b32_e32 v57, v2
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_mov_b32_e32 v58, v1
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_mov_b32_e32 v59, v0
|
|
; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v26
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v2, 16, v25
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v23
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v22
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v21
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v8, 16, v48
|
|
; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v49
|
|
; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v50
|
|
; SI-NEXT: v_lshrrev_b32_e32 v11, 16, v51
|
|
; SI-NEXT: v_lshrrev_b32_e32 v12, 16, v52
|
|
; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v53
|
|
; SI-NEXT: v_lshrrev_b32_e32 v14, 16, v54
|
|
; SI-NEXT: v_lshrrev_b32_e32 v15, 16, v55
|
|
; SI-NEXT: v_lshrrev_b32_e32 v16, 16, v40
|
|
; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v41
|
|
; SI-NEXT: v_lshrrev_b32_e32 v18, 16, v42
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v43
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v20, 16, v44
|
|
; SI-NEXT: v_lshrrev_b32_e32 v21, 16, v45
|
|
; SI-NEXT: v_lshrrev_b32_e32 v22, 16, v46
|
|
; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v47
|
|
; SI-NEXT: v_lshrrev_b32_e32 v24, 16, v56
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v57
|
|
; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v58
|
|
; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v59
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB50_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v55
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v59
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v58
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v57
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v56
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v47
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v46
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v45
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v44
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v43
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v42
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v41
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v39
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v38
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v62
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v37
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v32
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v36
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v61
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v35
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v60
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v34
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v63
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v33
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v54
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v53
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v52
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v51
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v50
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v49
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v27
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: .LBB50_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB50_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v55
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v59
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v58
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v57
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v56
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v47
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v46
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v45
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v44
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v43
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v42
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v41
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v40
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_or_b32_e32 v0, v39, v0
|
|
; SI-NEXT: s_mov_b32 s6, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v1, v38, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v62, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v37, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v32, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v36, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v61, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v35, v7
|
|
; SI-NEXT: v_or_b32_e32 v8, v60, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v34, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v63, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v33, v11
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, s6, v1
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v2
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, s6, v3
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v4
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, s6, v5
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v6
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, s6, v7
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v54
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v53
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v52
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v51
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v50
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v49
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v48
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v10
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, s6, v11
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v12
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, s6, v13
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v14
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, s6, v15
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v16
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, s6, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, s6, v19
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, s6, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, s6, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, s6, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 3, v25
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, s6, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 0x30000, v26
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 3, v27
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 0x30000, v27
|
|
; SI-NEXT: .LBB50_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56i16_to_v14f64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_mov_b32_e32 v32, v27
|
|
; VI-NEXT: v_mov_b32_e32 v33, v26
|
|
; VI-NEXT: v_mov_b32_e32 v34, v25
|
|
; VI-NEXT: v_mov_b32_e32 v35, v24
|
|
; VI-NEXT: v_mov_b32_e32 v36, v23
|
|
; VI-NEXT: v_mov_b32_e32 v37, v22
|
|
; VI-NEXT: v_mov_b32_e32 v38, v21
|
|
; VI-NEXT: v_mov_b32_e32 v39, v20
|
|
; VI-NEXT: v_mov_b32_e32 v48, v19
|
|
; VI-NEXT: v_mov_b32_e32 v49, v18
|
|
; VI-NEXT: v_mov_b32_e32 v50, v17
|
|
; VI-NEXT: v_mov_b32_e32 v51, v16
|
|
; VI-NEXT: v_mov_b32_e32 v52, v15
|
|
; VI-NEXT: v_mov_b32_e32 v53, v14
|
|
; VI-NEXT: v_mov_b32_e32 v54, v13
|
|
; VI-NEXT: v_mov_b32_e32 v55, v12
|
|
; VI-NEXT: v_mov_b32_e32 v40, v11
|
|
; VI-NEXT: v_mov_b32_e32 v41, v10
|
|
; VI-NEXT: v_mov_b32_e32 v42, v9
|
|
; VI-NEXT: v_mov_b32_e32 v43, v8
|
|
; VI-NEXT: v_mov_b32_e32 v44, v7
|
|
; VI-NEXT: v_mov_b32_e32 v45, v6
|
|
; VI-NEXT: v_mov_b32_e32 v46, v5
|
|
; VI-NEXT: v_mov_b32_e32 v47, v4
|
|
; VI-NEXT: v_mov_b32_e32 v56, v3
|
|
; VI-NEXT: v_mov_b32_e32 v57, v2
|
|
; VI-NEXT: v_mov_b32_e32 v58, v1
|
|
; VI-NEXT: v_mov_b32_e32 v59, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB50_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_mov_b32_e32 v27, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v27, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v27, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v27, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v27, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v27, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v27, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v27, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v27, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v27, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v27, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v27, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v27, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v27, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v27, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v27, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v27, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v27, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v27, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v27, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v27, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v27, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v24, v27, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v25, v27, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v26, v27, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v27, v27, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v56, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v47, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v46, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v45, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v44, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v43, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v42, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v41, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v40, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v53, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v52, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v51, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v50, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v49, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v39, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v37, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v36, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v35, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v34, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v33, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v32, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: ; implicit-def: $vgpr59
|
|
; VI-NEXT: ; implicit-def: $vgpr58
|
|
; VI-NEXT: ; implicit-def: $vgpr57
|
|
; VI-NEXT: ; implicit-def: $vgpr56
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: .LBB50_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB50_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 3
|
|
; VI-NEXT: v_add_u16_e32 v0, 3, v59
|
|
; VI-NEXT: v_add_u16_sdwa v1, v59, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v58
|
|
; VI-NEXT: v_add_u16_sdwa v3, v58, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v3
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v57
|
|
; VI-NEXT: v_add_u16_sdwa v3, v57, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; VI-NEXT: v_add_u16_e32 v3, 3, v56
|
|
; VI-NEXT: v_add_u16_sdwa v4, v56, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; VI-NEXT: v_add_u16_e32 v4, 3, v47
|
|
; VI-NEXT: v_add_u16_sdwa v5, v47, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; VI-NEXT: v_add_u16_e32 v5, 3, v46
|
|
; VI-NEXT: v_add_u16_sdwa v6, v46, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; VI-NEXT: v_add_u16_e32 v6, 3, v45
|
|
; VI-NEXT: v_add_u16_sdwa v7, v45, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; VI-NEXT: v_add_u16_e32 v7, 3, v44
|
|
; VI-NEXT: v_add_u16_sdwa v8, v44, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; VI-NEXT: v_add_u16_e32 v8, 3, v43
|
|
; VI-NEXT: v_add_u16_sdwa v9, v43, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; VI-NEXT: v_add_u16_e32 v9, 3, v42
|
|
; VI-NEXT: v_add_u16_sdwa v10, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; VI-NEXT: v_add_u16_e32 v10, 3, v41
|
|
; VI-NEXT: v_add_u16_sdwa v11, v41, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v10, v10, v11
|
|
; VI-NEXT: v_add_u16_e32 v11, 3, v40
|
|
; VI-NEXT: v_add_u16_sdwa v12, v40, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; VI-NEXT: v_add_u16_e32 v12, 3, v55
|
|
; VI-NEXT: v_add_u16_sdwa v13, v55, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; VI-NEXT: v_add_u16_e32 v13, 3, v54
|
|
; VI-NEXT: v_add_u16_sdwa v14, v54, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; VI-NEXT: v_add_u16_e32 v14, 3, v53
|
|
; VI-NEXT: v_add_u16_sdwa v15, v53, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; VI-NEXT: v_add_u16_e32 v15, 3, v52
|
|
; VI-NEXT: v_add_u16_sdwa v16, v52, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; VI-NEXT: v_add_u16_e32 v16, 3, v51
|
|
; VI-NEXT: v_add_u16_sdwa v17, v51, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; VI-NEXT: v_add_u16_e32 v17, 3, v50
|
|
; VI-NEXT: v_add_u16_sdwa v18, v50, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; VI-NEXT: v_add_u16_e32 v18, 3, v49
|
|
; VI-NEXT: v_add_u16_sdwa v19, v49, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; VI-NEXT: v_add_u16_e32 v19, 3, v48
|
|
; VI-NEXT: v_add_u16_sdwa v20, v48, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; VI-NEXT: v_add_u16_e32 v20, 3, v39
|
|
; VI-NEXT: v_add_u16_sdwa v21, v39, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; VI-NEXT: v_add_u16_e32 v21, 3, v38
|
|
; VI-NEXT: v_add_u16_sdwa v22, v38, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; VI-NEXT: v_add_u16_e32 v22, 3, v37
|
|
; VI-NEXT: v_add_u16_sdwa v23, v37, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; VI-NEXT: v_add_u16_e32 v23, 3, v36
|
|
; VI-NEXT: v_add_u16_sdwa v24, v36, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; VI-NEXT: v_add_u16_e32 v24, 3, v35
|
|
; VI-NEXT: v_add_u16_sdwa v25, v35, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; VI-NEXT: v_add_u16_e32 v25, 3, v34
|
|
; VI-NEXT: v_add_u16_sdwa v26, v34, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; VI-NEXT: v_add_u16_e32 v26, 3, v33
|
|
; VI-NEXT: v_add_u16_sdwa v28, v33, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; VI-NEXT: v_add_u16_e32 v28, 3, v32
|
|
; VI-NEXT: v_add_u16_sdwa v27, v32, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; VI-NEXT: .LBB50_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56i16_to_v14f64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_mov_b32_e32 v59, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v22
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v13
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v12
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v11
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v10
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v40
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v56, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v57, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v58, v1
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v41
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v47
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v56
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v57
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v58
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v59
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB50_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v37
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: .LBB50_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB50_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB50_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v14f64:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: .LBB50_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56i16_to_v14f64:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v69, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v71, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v80, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v68, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v67, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v66, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v65, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v64, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v54, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v53, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v52, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v51, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v50, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v49, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v48, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v39, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v38, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v37, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v36, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v35, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v34, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v33, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v32, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v31, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v30, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v29, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB50_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: .LBB50_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <56 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <56 x i16> %a1 to <14 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x i16> %a to <14 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x double> %phi
|
|
}
|
|
|
|
define inreg <14 x double> @bitcast_v56i16_to_v14f64_scalar(<56 x i16> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v56i16_to_v14f64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; SI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; SI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; SI-NEXT: v_writelane_b32 v28, s54, 14
|
|
; SI-NEXT: v_writelane_b32 v28, s55, 15
|
|
; SI-NEXT: v_writelane_b32 v28, s64, 16
|
|
; SI-NEXT: v_writelane_b32 v28, s65, 17
|
|
; SI-NEXT: v_writelane_b32 v28, s66, 18
|
|
; SI-NEXT: v_writelane_b32 v28, s67, 19
|
|
; SI-NEXT: v_writelane_b32 v28, s68, 20
|
|
; SI-NEXT: v_writelane_b32 v28, s69, 21
|
|
; SI-NEXT: v_writelane_b32 v28, s70, 22
|
|
; SI-NEXT: v_writelane_b32 v28, s71, 23
|
|
; SI-NEXT: v_writelane_b32 v28, s80, 24
|
|
; SI-NEXT: v_writelane_b32 v28, s81, 25
|
|
; SI-NEXT: v_writelane_b32 v28, s82, 26
|
|
; SI-NEXT: v_writelane_b32 v28, s83, 27
|
|
; SI-NEXT: v_writelane_b32 v28, s84, 28
|
|
; SI-NEXT: v_writelane_b32 v28, s85, 29
|
|
; SI-NEXT: v_writelane_b32 v28, s86, 30
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s73, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s75, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s77, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s88, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s94, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s31, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s69, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s80, v0
|
|
; SI-NEXT: v_writelane_b32 v28, s87, 31
|
|
; SI-NEXT: s_lshr_b32 s78, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s92, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s95, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s68, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s70, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s81, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s82, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s83, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s84, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s85, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s86, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s87, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s6, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s8, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s10, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s12, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s14, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s72, s73, 16
|
|
; SI-NEXT: s_lshr_b32 s74, s75, 16
|
|
; SI-NEXT: s_lshr_b32 s76, s77, 16
|
|
; SI-NEXT: s_lshr_b32 s79, s88, 16
|
|
; SI-NEXT: s_lshr_b32 s90, s91, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s94, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s31, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s69, 16
|
|
; SI-NEXT: s_lshr_b32 s71, s80, 16
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB51_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s36, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s37, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s38, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s39, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s40, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s41, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s42, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s43, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s44, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s34, 16
|
|
; SI-NEXT: s_or_b32 s45, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s95, 16
|
|
; SI-NEXT: s_or_b32 s46, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s92, 16
|
|
; SI-NEXT: s_or_b32 s47, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s48, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; SI-NEXT: s_or_b32 s49, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s80, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s50, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s51, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s31, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s52, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s94, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s53, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; SI-NEXT: s_or_b32 s54, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s55, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; SI-NEXT: s_or_b32 s56, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; SI-NEXT: s_or_b32 s57, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; SI-NEXT: s_or_b32 s58, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; SI-NEXT: s_or_b32 s59, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s12, 16
|
|
; SI-NEXT: s_or_b32 s60, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s10, 16
|
|
; SI-NEXT: s_or_b32 s61, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s8, 16
|
|
; SI-NEXT: s_or_b32 s62, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s6, 16
|
|
; SI-NEXT: s_or_b32 s63, s4, s5
|
|
; SI-NEXT: s_cbranch_execnz .LBB51_3
|
|
; SI-NEXT: .LBB51_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s17, s17, 3
|
|
; SI-NEXT: s_add_i32 s36, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: s_add_i32 s37, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s19, s19, 3
|
|
; SI-NEXT: s_add_i32 s38, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: s_add_i32 s39, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s21, s21, 3
|
|
; SI-NEXT: s_add_i32 s40, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: s_add_i32 s41, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s42, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s43, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s44, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s34, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s45, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s95, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s46, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s92, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s47, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: s_add_i32 s48, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s80, s80, 3
|
|
; SI-NEXT: s_add_i32 s49, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s80, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s69, s69, 3
|
|
; SI-NEXT: s_add_i32 s50, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s31, s31, 3
|
|
; SI-NEXT: s_add_i32 s51, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s31, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s94, s94, 3
|
|
; SI-NEXT: s_add_i32 s52, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s94, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s91, s91, 3
|
|
; SI-NEXT: s_add_i32 s53, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s88, s88, 3
|
|
; SI-NEXT: s_add_i32 s54, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s77, s77, 3
|
|
; SI-NEXT: s_add_i32 s55, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s75, s75, 3
|
|
; SI-NEXT: s_add_i32 s56, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s73, s73, 3
|
|
; SI-NEXT: s_add_i32 s57, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s15, s15, 3
|
|
; SI-NEXT: s_add_i32 s58, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s13, s13, 3
|
|
; SI-NEXT: s_add_i32 s59, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s12, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s11, s11, 3
|
|
; SI-NEXT: s_add_i32 s60, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s10, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s9, s9, 3
|
|
; SI-NEXT: s_add_i32 s61, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s8, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s7, s7, 3
|
|
; SI-NEXT: s_add_i32 s62, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s6, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s63, s4, 0x30000
|
|
; SI-NEXT: .LBB51_3: ; %end
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_readlane_b32 s87, v28, 31
|
|
; SI-NEXT: v_readlane_b32 s86, v28, 30
|
|
; SI-NEXT: v_readlane_b32 s85, v28, 29
|
|
; SI-NEXT: v_readlane_b32 s84, v28, 28
|
|
; SI-NEXT: v_readlane_b32 s83, v28, 27
|
|
; SI-NEXT: v_readlane_b32 s82, v28, 26
|
|
; SI-NEXT: v_readlane_b32 s81, v28, 25
|
|
; SI-NEXT: v_readlane_b32 s80, v28, 24
|
|
; SI-NEXT: v_readlane_b32 s71, v28, 23
|
|
; SI-NEXT: v_readlane_b32 s70, v28, 22
|
|
; SI-NEXT: v_readlane_b32 s69, v28, 21
|
|
; SI-NEXT: v_readlane_b32 s68, v28, 20
|
|
; SI-NEXT: v_readlane_b32 s67, v28, 19
|
|
; SI-NEXT: v_readlane_b32 s66, v28, 18
|
|
; SI-NEXT: v_readlane_b32 s65, v28, 17
|
|
; SI-NEXT: v_readlane_b32 s64, v28, 16
|
|
; SI-NEXT: v_readlane_b32 s55, v28, 15
|
|
; SI-NEXT: v_readlane_b32 s54, v28, 14
|
|
; SI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB51_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; SI-NEXT: s_branch .LBB51_2
|
|
;
|
|
; VI-LABEL: bitcast_v56i16_to_v14f64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; VI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; VI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; VI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; VI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; VI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; VI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; VI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; VI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; VI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; VI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; VI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; VI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; VI-NEXT: v_writelane_b32 v28, s54, 14
|
|
; VI-NEXT: v_writelane_b32 v28, s55, 15
|
|
; VI-NEXT: v_writelane_b32 v28, s64, 16
|
|
; VI-NEXT: v_writelane_b32 v28, s65, 17
|
|
; VI-NEXT: v_writelane_b32 v28, s66, 18
|
|
; VI-NEXT: v_writelane_b32 v28, s67, 19
|
|
; VI-NEXT: v_writelane_b32 v28, s68, 20
|
|
; VI-NEXT: v_writelane_b32 v28, s69, 21
|
|
; VI-NEXT: v_writelane_b32 v28, s70, 22
|
|
; VI-NEXT: v_writelane_b32 v28, s71, 23
|
|
; VI-NEXT: v_writelane_b32 v28, s80, 24
|
|
; VI-NEXT: v_writelane_b32 v28, s81, 25
|
|
; VI-NEXT: v_writelane_b32 v28, s82, 26
|
|
; VI-NEXT: v_writelane_b32 v28, s83, 27
|
|
; VI-NEXT: v_writelane_b32 v28, s84, 28
|
|
; VI-NEXT: v_writelane_b32 v28, s85, 29
|
|
; VI-NEXT: v_writelane_b32 v28, s86, 30
|
|
; VI-NEXT: v_writelane_b32 v28, s87, 31
|
|
; VI-NEXT: v_readfirstlane_b32 s86, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s87, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s73, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s77, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s88, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s34, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s69, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s81, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s84, v0
|
|
; VI-NEXT: s_lshr_b32 s79, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s67, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s70, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s80, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s82, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s85, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s7, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s11, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s13, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s15, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s10, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s64, s86, 16
|
|
; VI-NEXT: s_lshr_b32 s65, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s66, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s14, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s87, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s73, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s77, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s88, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s91, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s34, 16
|
|
; VI-NEXT: s_lshr_b32 s68, s69, 16
|
|
; VI-NEXT: s_lshr_b32 s71, s81, 16
|
|
; VI-NEXT: s_lshr_b32 s83, s84, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB51_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_or_b32 s36, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_or_b32 s38, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s10, 16
|
|
; VI-NEXT: s_or_b32 s39, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; VI-NEXT: s_or_b32 s37, s40, s41
|
|
; VI-NEXT: s_or_b32 s40, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_or_b32 s41, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s42, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s43, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; VI-NEXT: s_or_b32 s44, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s45, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s5, s67, 16
|
|
; VI-NEXT: s_or_b32 s46, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s47, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s48, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s49, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s84
|
|
; VI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; VI-NEXT: s_or_b32 s50, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s81
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s51, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s69
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s52, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s34
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s53, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s91
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s54, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s88
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s55, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s77
|
|
; VI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; VI-NEXT: s_or_b32 s56, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s73
|
|
; VI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; VI-NEXT: s_or_b32 s57, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; VI-NEXT: s_or_b32 s58, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s87
|
|
; VI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; VI-NEXT: s_or_b32 s59, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; VI-NEXT: s_or_b32 s60, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s5, s66, 16
|
|
; VI-NEXT: s_or_b32 s61, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s5, s65, 16
|
|
; VI-NEXT: s_or_b32 s62, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s86
|
|
; VI-NEXT: s_lshl_b32 s5, s64, 16
|
|
; VI-NEXT: s_or_b32 s63, s4, s5
|
|
; VI-NEXT: s_cbranch_execnz .LBB51_3
|
|
; VI-NEXT: .LBB51_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_and_b32 s5, s17, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s10, s10, 16
|
|
; VI-NEXT: s_or_b32 s5, s10, s5
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s36, s4, 0x30000
|
|
; VI-NEXT: s_add_i32 s37, s5, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s38, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s39, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s40, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s41, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s42, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s43, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s44, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s45, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s67, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s46, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s47, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s48, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s84, s84, 3
|
|
; VI-NEXT: s_add_i32 s49, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s84, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s81, s81, 3
|
|
; VI-NEXT: s_add_i32 s50, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s81, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s69, s69, 3
|
|
; VI-NEXT: s_add_i32 s51, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s69, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s34, s34, 3
|
|
; VI-NEXT: s_add_i32 s52, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s34, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s91, s91, 3
|
|
; VI-NEXT: s_add_i32 s53, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s91, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s88, s88, 3
|
|
; VI-NEXT: s_add_i32 s54, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s77, s77, 3
|
|
; VI-NEXT: s_add_i32 s55, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s77, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s75, s73, 3
|
|
; VI-NEXT: s_add_i32 s56, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s75, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s76, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s73, s12, 3
|
|
; VI-NEXT: s_add_i32 s57, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s73, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s74, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s15, s87, 3
|
|
; VI-NEXT: s_add_i32 s58, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s15, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s72, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s13, s8, 3
|
|
; VI-NEXT: s_add_i32 s59, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s13, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s14, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s11, s9, 3
|
|
; VI-NEXT: s_add_i32 s60, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s11, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s66, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s9, s6, 3
|
|
; VI-NEXT: s_add_i32 s61, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s9, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s65, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s7, s86, 3
|
|
; VI-NEXT: s_add_i32 s62, s4, 0x30000
|
|
; VI-NEXT: s_and_b32 s4, s7, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s64, 16
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_add_i32 s63, s4, 0x30000
|
|
; VI-NEXT: .LBB51_3: ; %end
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_readlane_b32 s87, v28, 31
|
|
; VI-NEXT: v_readlane_b32 s86, v28, 30
|
|
; VI-NEXT: v_readlane_b32 s85, v28, 29
|
|
; VI-NEXT: v_readlane_b32 s84, v28, 28
|
|
; VI-NEXT: v_readlane_b32 s83, v28, 27
|
|
; VI-NEXT: v_readlane_b32 s82, v28, 26
|
|
; VI-NEXT: v_readlane_b32 s81, v28, 25
|
|
; VI-NEXT: v_readlane_b32 s80, v28, 24
|
|
; VI-NEXT: v_readlane_b32 s71, v28, 23
|
|
; VI-NEXT: v_readlane_b32 s70, v28, 22
|
|
; VI-NEXT: v_readlane_b32 s69, v28, 21
|
|
; VI-NEXT: v_readlane_b32 s68, v28, 20
|
|
; VI-NEXT: v_readlane_b32 s67, v28, 19
|
|
; VI-NEXT: v_readlane_b32 s66, v28, 18
|
|
; VI-NEXT: v_readlane_b32 s65, v28, 17
|
|
; VI-NEXT: v_readlane_b32 s64, v28, 16
|
|
; VI-NEXT: v_readlane_b32 s55, v28, 15
|
|
; VI-NEXT: v_readlane_b32 s54, v28, 14
|
|
; VI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; VI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; VI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; VI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; VI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; VI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; VI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; VI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; VI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; VI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; VI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB51_4:
|
|
; VI-NEXT: ; implicit-def: $vgpr29 : SGPR spill to VGPR lane
|
|
; VI-NEXT: s_mov_b32 vcc_lo, s64
|
|
; VI-NEXT: v_writelane_b32 v29, s75, 0
|
|
; VI-NEXT: s_mov_b32 vcc_hi, s65
|
|
; VI-NEXT: s_mov_b32 s75, s73
|
|
; VI-NEXT: s_mov_b32 s73, s12
|
|
; VI-NEXT: v_writelane_b32 v29, s25, 1
|
|
; VI-NEXT: s_mov_b32 s25, s91
|
|
; VI-NEXT: s_mov_b32 s91, s26
|
|
; VI-NEXT: s_mov_b32 s26, s27
|
|
; VI-NEXT: s_mov_b32 s27, s88
|
|
; VI-NEXT: s_mov_b32 s88, s28
|
|
; VI-NEXT: s_mov_b32 s28, s77
|
|
; VI-NEXT: s_mov_b32 s77, s29
|
|
; VI-NEXT: s_mov_b32 s29, s15
|
|
; VI-NEXT: s_mov_b32 s15, s87
|
|
; VI-NEXT: s_mov_b32 s87, s85
|
|
; VI-NEXT: s_mov_b32 s85, s82
|
|
; VI-NEXT: s_mov_b32 s82, s71
|
|
; VI-NEXT: s_mov_b32 s71, s68
|
|
; VI-NEXT: s_mov_b32 s68, s35
|
|
; VI-NEXT: s_mov_b32 s35, s31
|
|
; VI-NEXT: s_mov_b32 s31, s30
|
|
; VI-NEXT: s_mov_b32 s30, s90
|
|
; VI-NEXT: s_mov_b32 s90, s89
|
|
; VI-NEXT: s_mov_b32 s89, s79
|
|
; VI-NEXT: s_mov_b32 s79, s78
|
|
; VI-NEXT: s_mov_b32 s78, s76
|
|
; VI-NEXT: s_mov_b32 s76, s74
|
|
; VI-NEXT: s_mov_b32 s74, s72
|
|
; VI-NEXT: s_mov_b32 s72, s14
|
|
; VI-NEXT: s_mov_b32 s14, s66
|
|
; VI-NEXT: s_mov_b32 s12, s10
|
|
; VI-NEXT: s_mov_b32 s10, s13
|
|
; VI-NEXT: s_mov_b32 s13, s8
|
|
; VI-NEXT: s_mov_b32 s8, s11
|
|
; VI-NEXT: s_mov_b32 s11, s9
|
|
; VI-NEXT: s_mov_b32 s9, s6
|
|
; VI-NEXT: s_mov_b32 s6, s7
|
|
; VI-NEXT: s_mov_b32 s7, s86
|
|
; VI-NEXT: s_mov_b32 s86, s83
|
|
; VI-NEXT: s_mov_b32 s83, s80
|
|
; VI-NEXT: s_mov_b32 s80, s70
|
|
; VI-NEXT: s_mov_b32 s70, s67
|
|
; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; VI-NEXT: s_mov_b32 s67, s70
|
|
; VI-NEXT: s_mov_b32 s70, s80
|
|
; VI-NEXT: s_mov_b32 s80, s83
|
|
; VI-NEXT: s_mov_b32 s83, s86
|
|
; VI-NEXT: s_mov_b32 s86, s7
|
|
; VI-NEXT: s_mov_b32 s7, s6
|
|
; VI-NEXT: s_mov_b32 s6, s9
|
|
; VI-NEXT: s_mov_b32 s9, s11
|
|
; VI-NEXT: s_mov_b32 s11, s8
|
|
; VI-NEXT: s_mov_b32 s8, s13
|
|
; VI-NEXT: s_mov_b32 s13, s10
|
|
; VI-NEXT: s_mov_b32 s10, s12
|
|
; VI-NEXT: s_mov_b32 s66, s14
|
|
; VI-NEXT: s_mov_b32 s14, s72
|
|
; VI-NEXT: s_mov_b32 s72, s74
|
|
; VI-NEXT: s_mov_b32 s74, s76
|
|
; VI-NEXT: s_mov_b32 s76, s78
|
|
; VI-NEXT: s_mov_b32 s78, s79
|
|
; VI-NEXT: s_mov_b32 s79, s89
|
|
; VI-NEXT: s_mov_b32 s89, s90
|
|
; VI-NEXT: s_mov_b32 s90, s30
|
|
; VI-NEXT: s_mov_b32 s30, s31
|
|
; VI-NEXT: s_mov_b32 s31, s35
|
|
; VI-NEXT: s_mov_b32 s35, s68
|
|
; VI-NEXT: s_mov_b32 s68, s71
|
|
; VI-NEXT: s_mov_b32 s71, s82
|
|
; VI-NEXT: s_mov_b32 s82, s85
|
|
; VI-NEXT: s_mov_b32 s85, s87
|
|
; VI-NEXT: s_mov_b32 s87, s15
|
|
; VI-NEXT: s_mov_b32 s15, s29
|
|
; VI-NEXT: s_mov_b32 s29, s77
|
|
; VI-NEXT: s_mov_b32 s77, s28
|
|
; VI-NEXT: s_mov_b32 s28, s88
|
|
; VI-NEXT: s_mov_b32 s88, s27
|
|
; VI-NEXT: s_mov_b32 s27, s26
|
|
; VI-NEXT: s_mov_b32 s26, s91
|
|
; VI-NEXT: s_mov_b32 s91, s25
|
|
; VI-NEXT: v_readlane_b32 s25, v29, 1
|
|
; VI-NEXT: s_mov_b32 s12, s73
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: v_readlane_b32 s75, v29, 0
|
|
; VI-NEXT: s_mov_b32 s65, vcc_hi
|
|
; VI-NEXT: s_mov_b32 s64, vcc_lo
|
|
; VI-NEXT: s_branch .LBB51_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v56i16_to_v14f64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: s_lshr_b32 s12, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s13, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s14, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s15, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s88, v5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v3
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v2
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: s_lshr_b32 s4, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s5, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s6, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s7, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s8, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s9, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s10, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s63, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s62, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s61, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s60, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s59, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s58, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s57, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s56, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s88, 16
|
|
; GFX9-NEXT: s_lshr_b32 s17, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s18, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s19, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s20, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s21, s12, 16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_cmp_lg_u32 s42, 0
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s50, s12, s21
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s51, s13, s20
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s52, s14, s19
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s53, s15, s18
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s54, s16, s17
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s55, s88, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s56, s56, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s57, s57, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s58, s58, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s59, s59, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s60, s60, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s61, s61, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s62, s62, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s63, s63, s72
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB51_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB51_4
|
|
; GFX9-NEXT: .LBB51_2: ; %cmp.true
|
|
; GFX9-NEXT: v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_branch .LBB51_5
|
|
; GFX9-NEXT: .LBB51_3:
|
|
; GFX9-NEXT: s_branch .LBB51_2
|
|
; GFX9-NEXT: .LBB51_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB51_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v56i16_to_v14f64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s44, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s74, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s76, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s77, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s78, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s88, v10
|
|
; GFX11-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s0, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s44, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s56, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s58, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s59, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s60, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s62, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s74, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s76, 16
|
|
; GFX11-NEXT: s_lshr_b32 s90, s77, 16
|
|
; GFX11-NEXT: s_lshr_b32 s91, s78, 16
|
|
; GFX11-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s88, 0
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s78, s91
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s77, s90
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s76, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s74, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s62, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s60, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s59, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s58, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s56, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s44, s57
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB51_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB51_4
|
|
; GFX11-NEXT: .LBB51_2: ; %cmp.true
|
|
; GFX11-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v18, s18, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v19, s19, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v20, s20, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v21, s21, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v22, s22, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v23, s23, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v24, s24, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v25, s25, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v26, s26, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: v_pk_add_u16 v27, s27, 3 op_sel_hi:[1,0]
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB51_3:
|
|
; GFX11-NEXT: s_branch .LBB51_2
|
|
; GFX11-NEXT: .LBB51_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <56 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <56 x i16> %a1 to <14 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x i16> %a to <14 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x double> %phi
|
|
}
|
|
|
|
define <56 x half> @bitcast_v14f64_to_v56f16(<14 x double> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v14f64_to_v56f16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB52_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v37, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v39, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB52_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB52_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; SI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; SI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; SI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; SI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; SI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; SI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; SI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; SI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; SI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; SI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; SI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; SI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; SI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; SI-NEXT: v_alignbit_b32 v28, v27, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v25, v24, 16
|
|
; SI-NEXT: v_alignbit_b32 v30, v23, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v21, v20, 16
|
|
; SI-NEXT: v_alignbit_b32 v32, v19, v18, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v17, v16, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v15, v14, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v13, v12, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v11, v10, 16
|
|
; SI-NEXT: v_alignbit_b32 v37, v9, v8, 16
|
|
; SI-NEXT: v_alignbit_b32 v39, v7, v6, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v5, v4, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v3, v2, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v1, v0, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: .LBB52_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v47
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v40
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v41
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v53
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v50
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v39
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v37
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14f64_to_v56f16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB52_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB52_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB52_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; VI-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; VI-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; VI-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; VI-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; VI-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; VI-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; VI-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; VI-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; VI-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; VI-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; VI-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; VI-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; VI-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; VI-NEXT: .LBB52_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14f64_to_v56f16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB52_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB52_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB52_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: .LBB52_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v47, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v46, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v45, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v44, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v43, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v42, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v41, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v40, v7, s4
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v8, v55, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v54, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v53, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v52, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v51, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v49, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v48, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v39, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v38, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v37, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v36, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v35, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v34, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v33, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v32, v23, s4
|
|
; GFX9-NEXT: v_perm_b32 v24, v31, v24, s4
|
|
; GFX9-NEXT: v_perm_b32 v25, v30, v25, s4
|
|
; GFX9-NEXT: v_perm_b32 v26, v29, v26, s4
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s4
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v14f64_to_v56f16:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB52_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX11-TRUE16-NEXT: .LBB52_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v14f64_to_v56f16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr71
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr70
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr69
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr68
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr67
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr66
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr65
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr64
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB52_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB52_2: ; %Flow
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB52_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[26:27], v[26:27], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[16:17], v[16:17], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[14:15], v[14:15], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[12:13], v[12:13], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[10:11], v[10:11], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[8:9], v[8:9], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[6:7], v[6:7], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[4:5], v[4:5], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[2:3], v[2:3], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[0:1], v[0:1], 1.0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-FAKE16-NEXT: .LBB52_4: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v71, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v69, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v68, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v67, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v66, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v65, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v64, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v55, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v54, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v53, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v52, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v51, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v49, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v39, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v38, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v37, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v36, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v35, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v34, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v33, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v32, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v31, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v30, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v29, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v28, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <14 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <14 x double> %a1 to <56 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x double> %a to <56 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x half> %phi
|
|
}
|
|
|
|
define inreg <56 x half> @bitcast_v14f64_to_v56f16_scalar(<14 x double> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v14f64_to_v56f16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v58, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v58, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v58, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v58, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v58, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v58, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v58, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v58, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v58, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v58, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v58, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v58, s51, 11
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: v_writelane_b32 v58, s52, 12
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s5, v1
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v0
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_writelane_b32 v58, s53, 13
|
|
; SI-NEXT: s_cbranch_scc0 .LBB53_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s53, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s52, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s48, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s38, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s36, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s17, 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[28:29], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[26:27], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[88:89], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[92:93], s[16:17], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB53_4
|
|
; SI-NEXT: .LBB53_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[26:27], s[40:41], 1.0
|
|
; SI-NEXT: v_add_f64 v[24:25], s[14:15], 1.0
|
|
; SI-NEXT: v_add_f64 v[22:23], s[12:13], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[28:29], v[26:27], 16
|
|
; SI-NEXT: v_add_f64 v[20:21], s[10:11], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[29:30], v[24:25], 16
|
|
; SI-NEXT: v_add_f64 v[18:19], s[8:9], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[30:31], v[22:23], 16
|
|
; SI-NEXT: v_add_f64 v[16:17], s[6:7], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[31:32], v[20:21], 16
|
|
; SI-NEXT: v_add_f64 v[14:15], s[4:5], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[32:33], v[18:19], 16
|
|
; SI-NEXT: v_add_f64 v[12:13], s[28:29], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[33:34], v[16:17], 16
|
|
; SI-NEXT: v_add_f64 v[10:11], s[26:27], 1.0
|
|
; SI-NEXT: v_add_f64 v[4:5], s[20:21], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[34:35], v[14:15], 16
|
|
; SI-NEXT: v_add_f64 v[8:9], s[24:25], 1.0
|
|
; SI-NEXT: v_add_f64 v[2:3], s[18:19], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[35:36], v[12:13], 16
|
|
; SI-NEXT: v_add_f64 v[6:7], s[22:23], 1.0
|
|
; SI-NEXT: v_add_f64 v[0:1], s[16:17], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[36:37], v[10:11], 16
|
|
; SI-NEXT: v_lshr_b64 v[48:49], v[4:5], 16
|
|
; SI-NEXT: v_lshr_b64 v[37:38], v[8:9], 16
|
|
; SI-NEXT: v_lshr_b64 v[49:50], v[2:3], 16
|
|
; SI-NEXT: v_lshr_b64 v[38:39], v[6:7], 16
|
|
; SI-NEXT: v_lshr_b64 v[50:51], v[0:1], 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v15
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v13
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v11
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v9
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v1
|
|
; SI-NEXT: s_branch .LBB53_5
|
|
; SI-NEXT: .LBB53_3:
|
|
; SI-NEXT: ; implicit-def: $sgpr92
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: ; implicit-def: $sgpr31
|
|
; SI-NEXT: ; implicit-def: $sgpr88
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr78
|
|
; SI-NEXT: ; implicit-def: $sgpr35
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr36
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr37
|
|
; SI-NEXT: ; implicit-def: $sgpr72
|
|
; SI-NEXT: ; implicit-def: $sgpr38
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr39
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr48
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr49
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr50
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr51
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr52
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr53
|
|
; SI-NEXT: s_branch .LBB53_2
|
|
; SI-NEXT: .LBB53_4:
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v0, s16
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v4, s20
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v6, s22
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v8, s24
|
|
; SI-NEXT: v_mov_b32_e32 v27, s41
|
|
; SI-NEXT: v_mov_b32_e32 v25, s15
|
|
; SI-NEXT: v_mov_b32_e32 v23, s13
|
|
; SI-NEXT: v_mov_b32_e32 v21, s11
|
|
; SI-NEXT: v_mov_b32_e32 v19, s9
|
|
; SI-NEXT: v_mov_b32_e32 v17, s7
|
|
; SI-NEXT: v_mov_b32_e32 v15, s5
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v10, s26
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v14, s4
|
|
; SI-NEXT: v_mov_b32_e32 v16, s6
|
|
; SI-NEXT: v_mov_b32_e32 v18, s8
|
|
; SI-NEXT: v_mov_b32_e32 v20, s10
|
|
; SI-NEXT: v_mov_b32_e32 v22, s12
|
|
; SI-NEXT: v_mov_b32_e32 v24, s14
|
|
; SI-NEXT: v_mov_b32_e32 v26, s40
|
|
; SI-NEXT: v_mov_b32_e32 v52, s53
|
|
; SI-NEXT: v_mov_b32_e32 v53, s52
|
|
; SI-NEXT: v_mov_b32_e32 v54, s51
|
|
; SI-NEXT: v_mov_b32_e32 v55, s50
|
|
; SI-NEXT: v_mov_b32_e32 v40, s49
|
|
; SI-NEXT: v_mov_b32_e32 v41, s48
|
|
; SI-NEXT: v_mov_b32_e32 v42, s39
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_mov_b32_e32 v43, s38
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_mov_b32_e32 v44, s37
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_mov_b32_e32 v45, s36
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_mov_b32_e32 v46, s35
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v47, s34
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_mov_b32_e32 v56, s31
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v57, s30
|
|
; SI-NEXT: v_mov_b32_e32 v50, s92
|
|
; SI-NEXT: v_mov_b32_e32 v49, s90
|
|
; SI-NEXT: v_mov_b32_e32 v48, s88
|
|
; SI-NEXT: v_mov_b32_e32 v38, s78
|
|
; SI-NEXT: v_mov_b32_e32 v37, s76
|
|
; SI-NEXT: v_mov_b32_e32 v36, s74
|
|
; SI-NEXT: v_mov_b32_e32 v35, s72
|
|
; SI-NEXT: v_mov_b32_e32 v34, s62
|
|
; SI-NEXT: v_mov_b32_e32 v33, s60
|
|
; SI-NEXT: v_mov_b32_e32 v32, s58
|
|
; SI-NEXT: v_mov_b32_e32 v31, s56
|
|
; SI-NEXT: v_mov_b32_e32 v30, s46
|
|
; SI-NEXT: v_mov_b32_e32 v29, s44
|
|
; SI-NEXT: v_mov_b32_e32 v28, s42
|
|
; SI-NEXT: .LBB53_5: ; %end
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v50
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v39
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v57
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v49
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v39
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v56
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v48
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v34
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v33
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v41
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v40
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v31
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v30
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v29
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v39
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v38
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v37
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: v_readlane_b32 s53, v58, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v58, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v58, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v58, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v58, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v58, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v58, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v58, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v58, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v58, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v58, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v58, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v58, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v58, 0
|
|
; SI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v14f64_to_v56f16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v56, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v56, s31, 1
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v56, s34, 2
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s5, v1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v0
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_writelane_b32 v56, s35, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB53_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s44, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s5, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s4, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB53_4
|
|
; VI-NEXT: .LBB53_2: ; %cmp.true
|
|
; VI-NEXT: v_add_f64 v[26:27], s[8:9], 1.0
|
|
; VI-NEXT: v_add_f64 v[24:25], s[10:11], 1.0
|
|
; VI-NEXT: v_add_f64 v[22:23], s[12:13], 1.0
|
|
; VI-NEXT: v_add_f64 v[20:21], s[14:15], 1.0
|
|
; VI-NEXT: v_add_f64 v[18:19], s[40:41], 1.0
|
|
; VI-NEXT: v_add_f64 v[16:17], s[6:7], 1.0
|
|
; VI-NEXT: v_add_f64 v[14:15], s[4:5], 1.0
|
|
; VI-NEXT: v_add_f64 v[12:13], s[28:29], 1.0
|
|
; VI-NEXT: v_add_f64 v[10:11], s[26:27], 1.0
|
|
; VI-NEXT: v_add_f64 v[8:9], s[24:25], 1.0
|
|
; VI-NEXT: v_add_f64 v[6:7], s[22:23], 1.0
|
|
; VI-NEXT: v_add_f64 v[4:5], s[20:21], 1.0
|
|
; VI-NEXT: v_add_f64 v[2:3], s[18:19], 1.0
|
|
; VI-NEXT: v_add_f64 v[0:1], s[16:17], 1.0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v0
|
|
; VI-NEXT: s_branch .LBB53_5
|
|
; VI-NEXT: .LBB53_3:
|
|
; VI-NEXT: ; implicit-def: $sgpr35
|
|
; VI-NEXT: ; implicit-def: $sgpr73
|
|
; VI-NEXT: ; implicit-def: $sgpr34
|
|
; VI-NEXT: ; implicit-def: $sgpr72
|
|
; VI-NEXT: ; implicit-def: $sgpr31
|
|
; VI-NEXT: ; implicit-def: $sgpr63
|
|
; VI-NEXT: ; implicit-def: $sgpr30
|
|
; VI-NEXT: ; implicit-def: $sgpr62
|
|
; VI-NEXT: ; implicit-def: $sgpr91
|
|
; VI-NEXT: ; implicit-def: $sgpr61
|
|
; VI-NEXT: ; implicit-def: $sgpr90
|
|
; VI-NEXT: ; implicit-def: $sgpr60
|
|
; VI-NEXT: ; implicit-def: $sgpr89
|
|
; VI-NEXT: ; implicit-def: $sgpr59
|
|
; VI-NEXT: ; implicit-def: $sgpr88
|
|
; VI-NEXT: ; implicit-def: $sgpr58
|
|
; VI-NEXT: ; implicit-def: $sgpr79
|
|
; VI-NEXT: ; implicit-def: $sgpr57
|
|
; VI-NEXT: ; implicit-def: $sgpr78
|
|
; VI-NEXT: ; implicit-def: $sgpr56
|
|
; VI-NEXT: ; implicit-def: $sgpr77
|
|
; VI-NEXT: ; implicit-def: $sgpr47
|
|
; VI-NEXT: ; implicit-def: $sgpr76
|
|
; VI-NEXT: ; implicit-def: $sgpr46
|
|
; VI-NEXT: ; implicit-def: $sgpr75
|
|
; VI-NEXT: ; implicit-def: $sgpr45
|
|
; VI-NEXT: ; implicit-def: $sgpr74
|
|
; VI-NEXT: ; implicit-def: $sgpr44
|
|
; VI-NEXT: s_branch .LBB53_2
|
|
; VI-NEXT: .LBB53_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v14, s4
|
|
; VI-NEXT: v_mov_b32_e32 v16, s6
|
|
; VI-NEXT: v_mov_b32_e32 v18, s40
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v15, s5
|
|
; VI-NEXT: v_mov_b32_e32 v17, s7
|
|
; VI-NEXT: v_mov_b32_e32 v19, s41
|
|
; VI-NEXT: v_mov_b32_e32 v20, s14
|
|
; VI-NEXT: v_mov_b32_e32 v21, s15
|
|
; VI-NEXT: v_mov_b32_e32 v22, s12
|
|
; VI-NEXT: v_mov_b32_e32 v23, s13
|
|
; VI-NEXT: v_mov_b32_e32 v24, s10
|
|
; VI-NEXT: v_mov_b32_e32 v25, s11
|
|
; VI-NEXT: v_mov_b32_e32 v26, s8
|
|
; VI-NEXT: v_mov_b32_e32 v27, s9
|
|
; VI-NEXT: v_mov_b32_e32 v44, s35
|
|
; VI-NEXT: v_mov_b32_e32 v42, s34
|
|
; VI-NEXT: v_mov_b32_e32 v40, s31
|
|
; VI-NEXT: v_mov_b32_e32 v54, s30
|
|
; VI-NEXT: v_mov_b32_e32 v52, s91
|
|
; VI-NEXT: v_mov_b32_e32 v50, s90
|
|
; VI-NEXT: v_mov_b32_e32 v48, s89
|
|
; VI-NEXT: v_mov_b32_e32 v38, s88
|
|
; VI-NEXT: v_mov_b32_e32 v36, s79
|
|
; VI-NEXT: v_mov_b32_e32 v34, s78
|
|
; VI-NEXT: v_mov_b32_e32 v32, s77
|
|
; VI-NEXT: v_mov_b32_e32 v30, s76
|
|
; VI-NEXT: v_mov_b32_e32 v29, s75
|
|
; VI-NEXT: v_mov_b32_e32 v28, s74
|
|
; VI-NEXT: v_mov_b32_e32 v47, s73
|
|
; VI-NEXT: v_mov_b32_e32 v46, s72
|
|
; VI-NEXT: v_mov_b32_e32 v45, s63
|
|
; VI-NEXT: v_mov_b32_e32 v43, s62
|
|
; VI-NEXT: v_mov_b32_e32 v41, s61
|
|
; VI-NEXT: v_mov_b32_e32 v55, s60
|
|
; VI-NEXT: v_mov_b32_e32 v53, s59
|
|
; VI-NEXT: v_mov_b32_e32 v51, s58
|
|
; VI-NEXT: v_mov_b32_e32 v49, s57
|
|
; VI-NEXT: v_mov_b32_e32 v39, s56
|
|
; VI-NEXT: v_mov_b32_e32 v37, s47
|
|
; VI-NEXT: v_mov_b32_e32 v35, s46
|
|
; VI-NEXT: v_mov_b32_e32 v33, s45
|
|
; VI-NEXT: v_mov_b32_e32 v31, s44
|
|
; VI-NEXT: .LBB53_5: ; %end
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v47
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v46
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v43
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v41
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v55
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v53
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v51
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v49
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v39
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v37
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v35
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v33
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v31
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_readlane_b32 s35, v56, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v56, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v56, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v56, 0
|
|
; VI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v14f64_to_v56f16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s41, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s40, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s5, v1
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v0
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB53_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s44, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s41, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s40, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s6, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s5, 16
|
|
; GFX9-NEXT: s_lshr_b32 s88, s4, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s90, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s91, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s92, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s93, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s94, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s95, s16, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB53_4
|
|
; GFX9-NEXT: .LBB53_2: ; %cmp.true
|
|
; GFX9-NEXT: v_add_f64 v[26:27], s[8:9], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[24:25], s[10:11], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[22:23], s[12:13], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[20:21], s[14:15], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[18:19], s[40:41], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[16:17], s[6:7], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[14:15], s[4:5], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[12:13], s[28:29], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[10:11], s[26:27], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[8:9], s[24:25], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[6:7], s[22:23], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[4:5], s[20:21], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[2:3], s[18:19], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[0:1], s[16:17], 1.0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v0
|
|
; GFX9-NEXT: s_branch .LBB53_5
|
|
; GFX9-NEXT: .LBB53_3:
|
|
; GFX9-NEXT: ; implicit-def: $sgpr95
|
|
; GFX9-NEXT: ; implicit-def: $sgpr73
|
|
; GFX9-NEXT: ; implicit-def: $sgpr94
|
|
; GFX9-NEXT: ; implicit-def: $sgpr72
|
|
; GFX9-NEXT: ; implicit-def: $sgpr93
|
|
; GFX9-NEXT: ; implicit-def: $sgpr63
|
|
; GFX9-NEXT: ; implicit-def: $sgpr92
|
|
; GFX9-NEXT: ; implicit-def: $sgpr62
|
|
; GFX9-NEXT: ; implicit-def: $sgpr91
|
|
; GFX9-NEXT: ; implicit-def: $sgpr61
|
|
; GFX9-NEXT: ; implicit-def: $sgpr90
|
|
; GFX9-NEXT: ; implicit-def: $sgpr60
|
|
; GFX9-NEXT: ; implicit-def: $sgpr89
|
|
; GFX9-NEXT: ; implicit-def: $sgpr59
|
|
; GFX9-NEXT: ; implicit-def: $sgpr88
|
|
; GFX9-NEXT: ; implicit-def: $sgpr58
|
|
; GFX9-NEXT: ; implicit-def: $sgpr79
|
|
; GFX9-NEXT: ; implicit-def: $sgpr57
|
|
; GFX9-NEXT: ; implicit-def: $sgpr78
|
|
; GFX9-NEXT: ; implicit-def: $sgpr56
|
|
; GFX9-NEXT: ; implicit-def: $sgpr77
|
|
; GFX9-NEXT: ; implicit-def: $sgpr47
|
|
; GFX9-NEXT: ; implicit-def: $sgpr76
|
|
; GFX9-NEXT: ; implicit-def: $sgpr46
|
|
; GFX9-NEXT: ; implicit-def: $sgpr75
|
|
; GFX9-NEXT: ; implicit-def: $sgpr45
|
|
; GFX9-NEXT: ; implicit-def: $sgpr74
|
|
; GFX9-NEXT: ; implicit-def: $sgpr44
|
|
; GFX9-NEXT: s_branch .LBB53_2
|
|
; GFX9-NEXT: .LBB53_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s4
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s5
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, s95
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, s94
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, s93
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, s92
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, s91
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, s90
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, s89
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, s88
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, s79
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, s78
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, s77
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s76
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s75
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s74
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, s73
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, s72
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s44
|
|
; GFX9-NEXT: .LBB53_5: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v0, v44, 16, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v1, v47, 16, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v2, v42, 16, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v3, v46, 16, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v4, v40, 16, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v5, v45, 16, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v7, v43, 16, v7
|
|
; GFX9-NEXT: v_lshl_or_b32 v9, v41, 16, v9
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX9-NEXT: v_lshl_or_b32 v6, v54, 16, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v8, v52, 16, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v50, 16, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v55, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v48, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v53, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v38, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v51, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v36, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v49, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v34, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v39, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v32, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v37, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v30, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v35, 16, v23
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v29, 16, v24
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v33, 16, v25
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v28, 16, v26
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v31, 16, v27
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v14f64_to_v56f16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s11, v9
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s10, v8
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s9, v7
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s8, v6
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s7, v5
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s6, v4
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v3
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v2
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB53_3
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s11, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s61, s10, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s9, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s62, s8, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s63, s6, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s5, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s72, s4, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s13, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s74, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s75, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s76, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s77, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s78, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s57, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s79, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s58, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s88, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s59, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s89, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s60, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s90, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB53_4
|
|
; GFX11-TRUE16-NEXT: .LBB53_2: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[26:27], s[10:11], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[24:25], s[8:9], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[22:23], s[6:7], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[20:21], s[4:5], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[18:19], s[12:13], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[16:17], s[28:29], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[14:15], s[26:27], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[12:13], s[24:25], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[10:11], s[22:23], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[8:9], s[20:21], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[6:7], s[18:19], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[4:5], s[16:17], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[2:3], s[2:3], 1.0
|
|
; GFX11-TRUE16-NEXT: v_add_f64 v[0:1], s[0:1], 1.0
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v27
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v26
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v25
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v24
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v65, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v67, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 16, v0
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB53_5
|
|
; GFX11-TRUE16-NEXT: .LBB53_3:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr90
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB53_2
|
|
; GFX11-TRUE16-NEXT: .LBB53_4:
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, s12 :: v_dual_mov_b32 v19, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s4 :: v_dual_mov_b32 v21, s5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, s6 :: v_dual_mov_b32 v23, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v24, s8 :: v_dual_mov_b32 v25, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, s10 :: v_dual_mov_b32 v27, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v68, s90 :: v_dual_mov_b32 v29, s62
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v66, s89 :: v_dual_mov_b32 v71, s60
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, s88 :: v_dual_mov_b32 v69, s58
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s79 :: v_dual_mov_b32 v67, s57
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v52, s78 :: v_dual_mov_b32 v65, s56
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v50, s77 :: v_dual_mov_b32 v55, s47
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v48, s76 :: v_dual_mov_b32 v53, s46
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v38, s75 :: v_dual_mov_b32 v51, s45
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v36, s74 :: v_dual_mov_b32 v49, s44
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v34, s73 :: v_dual_mov_b32 v39, s43
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v32, s72 :: v_dual_mov_b32 v37, s42
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, s63 :: v_dual_mov_b32 v35, s41
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, s61 :: v_dual_mov_b32 v33, s40
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v70, s59 :: v_dual_mov_b32 v31, s15
|
|
; GFX11-TRUE16-NEXT: .LBB53_5: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v68.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v71.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v66.l
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v70.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v64.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v69.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v67.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v65.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v28.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v31.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v14f64_to_v56f16_scalar:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s14, v10
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s11, v9
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s10, v8
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s9, v7
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s8, v6
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s7, v5
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s6, v4
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s5, v3
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s4, v2
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s14, 0
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s14, 0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB53_3
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s11, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s61, s10, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s9, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s62, s8, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s7, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s63, s6, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s5, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s72, s4, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s13, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s74, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s75, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s76, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s77, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s78, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s57, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s79, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s58, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s88, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s59, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s89, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s60, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s90, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s14
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB53_4
|
|
; GFX11-FAKE16-NEXT: .LBB53_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[23:24], s[10:11], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[25:26], s[8:9], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[27:28], s[6:7], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[18:19], s[4:5], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[20:21], s[12:13], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[13:14], s[28:29], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[15:16], s[26:27], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[29:30], s[24:25], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[8:9], s[22:23], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[10:11], s[20:21], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[3:4], s[18:19], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[5:6], s[16:17], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[31:32], s[2:3], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[0:1], s[0:1], 1.0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v28
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v17, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v30
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v12, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v7, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v32
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v2, 16, v31
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB53_5
|
|
; GFX11-FAKE16-NEXT: .LBB53_3:
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr90
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr60
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr89
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr59
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr88
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr58
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr79
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr57
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr78
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr56
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr77
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr47
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr76
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr46
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr75
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr45
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr74
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr44
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr73
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr43
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr72
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr42
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr63
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr41
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr62
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr40
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr61
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $sgpr15
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB53_2
|
|
; GFX11-FAKE16-NEXT: .LBB53_4:
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v31, s2
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v10, s20
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v8, s22
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v29, s24 :: v_dual_mov_b32 v20, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v18, s4
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s28 :: v_dual_mov_b32 v32, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v27, s6 :: v_dual_mov_b32 v6, s17
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v25, s8 :: v_dual_mov_b32 v4, s19
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v23, s10 :: v_dual_mov_b32 v30, s25
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v16, s27
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v14, s29
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v28, s7
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v21, s13 :: v_dual_mov_b32 v26, s9
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v19, s5 :: v_dual_mov_b32 v24, s11
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v69, s90 :: v_dual_mov_b32 v2, s89
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v66, s88 :: v_dual_mov_b32 v7, s79
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v55, s78 :: v_dual_mov_b32 v12, s76
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v53, s77 :: v_dual_mov_b32 v50, s75
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s74 :: v_dual_mov_b32 v22, s63
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v39, s73 :: v_dual_mov_b32 v34, s62
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v37, s72 :: v_dual_mov_b32 v70, s59
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v33, s61 :: v_dual_mov_b32 v68, s58
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v71, s60 :: v_dual_mov_b32 v64, s47
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v67, s57 :: v_dual_mov_b32 v54, s46
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v65, s56 :: v_dual_mov_b32 v52, s45
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v51, s44 :: v_dual_mov_b32 v48, s42
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v49, s43 :: v_dual_mov_b32 v38, s41
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, s40 :: v_dual_mov_b32 v35, s15
|
|
; GFX11-FAKE16-NEXT: .LBB53_5: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v31
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v2, 16, v31
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v32
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v12, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v30
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v70, 16, v31
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xffff, v13
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v54, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v69, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v68, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v7, 16, v32
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v32, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v55, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v53, 16, v31
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v31, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v50, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v52, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v17, 16, v30
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v30, 0xffff, v19
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v39, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v37, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v22, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xffff, v28
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v28, 0xffff, v23
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v71, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v67, 16, v69
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v65, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v64, 16, v32
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v51, 16, v31
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v49, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v48, 16, v30
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v38, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v34, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v36, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v33, 16, v28
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v35, 16, v29
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <14 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <14 x double> %a1 to <56 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <14 x double> %a to <56 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x half> %phi
|
|
}
|
|
|
|
define <14 x double> @bitcast_v56f16_to_v14f64(<56 x half> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v56f16_to_v14f64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_mov_b32_e32 v59, v0
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v22
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v21
|
|
; SI-NEXT: v_mov_b32_e32 v48, v19
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v20
|
|
; SI-NEXT: v_mov_b32_e32 v49, v18
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; SI-NEXT: v_mov_b32_e32 v50, v17
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; SI-NEXT: v_mov_b32_e32 v51, v16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; SI-NEXT: v_mov_b32_e32 v52, v15
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; SI-NEXT: v_mov_b32_e32 v53, v14
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; SI-NEXT: v_mov_b32_e32 v54, v13
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; SI-NEXT: v_mov_b32_e32 v55, v12
|
|
; SI-NEXT: v_mov_b32_e32 v40, v11
|
|
; SI-NEXT: v_mov_b32_e32 v41, v10
|
|
; SI-NEXT: v_mov_b32_e32 v42, v9
|
|
; SI-NEXT: v_mov_b32_e32 v43, v8
|
|
; SI-NEXT: v_mov_b32_e32 v44, v7
|
|
; SI-NEXT: v_mov_b32_e32 v45, v6
|
|
; SI-NEXT: v_mov_b32_e32 v46, v5
|
|
; SI-NEXT: v_mov_b32_e32 v47, v4
|
|
; SI-NEXT: v_mov_b32_e32 v56, v3
|
|
; SI-NEXT: v_mov_b32_e32 v57, v2
|
|
; SI-NEXT: v_mov_b32_e32 v58, v1
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v40
|
|
; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v41
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v42
|
|
; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v43
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v44
|
|
; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v46
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v47
|
|
; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v56
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v57
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v58
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v59
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB54_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v59
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v39
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v58
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v3
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v57
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v62
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v47
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v46
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v45
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v61
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v43
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v60
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v41
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v63
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v11
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v40
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v55
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v53
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v52
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v51
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v50
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v49
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v27
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: .LBB54_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB54_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v32
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v58
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v56
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v47
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v62
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v46
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v45
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v37
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v44
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v43
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v36
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v42
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v41
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v61
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v40
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v55
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; SI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v35
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v54
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v53
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v60
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v52
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v51
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v34
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v49
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v48
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v63
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v33
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v29, 0x38000000, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v16
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v18, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v19
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v21, v19
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v25
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v27, v25
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_or_b32_e32 v24, v26, v24
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v27, v29, v27
|
|
; SI-NEXT: .LBB54_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56f16_to_v14f64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_mov_b32_e32 v32, v27
|
|
; VI-NEXT: v_mov_b32_e32 v33, v26
|
|
; VI-NEXT: v_mov_b32_e32 v34, v25
|
|
; VI-NEXT: v_mov_b32_e32 v35, v24
|
|
; VI-NEXT: v_mov_b32_e32 v36, v23
|
|
; VI-NEXT: v_mov_b32_e32 v37, v22
|
|
; VI-NEXT: v_mov_b32_e32 v38, v21
|
|
; VI-NEXT: v_mov_b32_e32 v39, v20
|
|
; VI-NEXT: v_mov_b32_e32 v48, v19
|
|
; VI-NEXT: v_mov_b32_e32 v49, v18
|
|
; VI-NEXT: v_mov_b32_e32 v50, v17
|
|
; VI-NEXT: v_mov_b32_e32 v51, v16
|
|
; VI-NEXT: v_mov_b32_e32 v52, v15
|
|
; VI-NEXT: v_mov_b32_e32 v53, v14
|
|
; VI-NEXT: v_mov_b32_e32 v54, v13
|
|
; VI-NEXT: v_mov_b32_e32 v55, v12
|
|
; VI-NEXT: v_mov_b32_e32 v40, v11
|
|
; VI-NEXT: v_mov_b32_e32 v41, v10
|
|
; VI-NEXT: v_mov_b32_e32 v42, v9
|
|
; VI-NEXT: v_mov_b32_e32 v43, v8
|
|
; VI-NEXT: v_mov_b32_e32 v44, v7
|
|
; VI-NEXT: v_mov_b32_e32 v45, v6
|
|
; VI-NEXT: v_mov_b32_e32 v46, v5
|
|
; VI-NEXT: v_mov_b32_e32 v47, v4
|
|
; VI-NEXT: v_mov_b32_e32 v56, v3
|
|
; VI-NEXT: v_mov_b32_e32 v57, v2
|
|
; VI-NEXT: v_mov_b32_e32 v58, v1
|
|
; VI-NEXT: v_mov_b32_e32 v59, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB54_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_mov_b32_e32 v27, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v27, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v27, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v27, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v27, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v27, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v27, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v27, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v27, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v27, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v27, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v27, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v27, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v27, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v27, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v27, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v27, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v27, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v27, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v27, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v27, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v27, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v24, v27, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v25, v27, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v26, v27, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v27, v27, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v58, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v57, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v56, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v47, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v46, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v45, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v44, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v43, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v42, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v41, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v40, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v53, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v52, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v51, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v50, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v49, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v48, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v39, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v37, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v36, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v35, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v34, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v33, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v32, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: ; implicit-def: $vgpr59
|
|
; VI-NEXT: ; implicit-def: $vgpr58
|
|
; VI-NEXT: ; implicit-def: $vgpr57
|
|
; VI-NEXT: ; implicit-def: $vgpr56
|
|
; VI-NEXT: ; implicit-def: $vgpr47
|
|
; VI-NEXT: ; implicit-def: $vgpr46
|
|
; VI-NEXT: ; implicit-def: $vgpr45
|
|
; VI-NEXT: ; implicit-def: $vgpr44
|
|
; VI-NEXT: ; implicit-def: $vgpr43
|
|
; VI-NEXT: ; implicit-def: $vgpr42
|
|
; VI-NEXT: ; implicit-def: $vgpr41
|
|
; VI-NEXT: ; implicit-def: $vgpr40
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr54
|
|
; VI-NEXT: ; implicit-def: $vgpr53
|
|
; VI-NEXT: ; implicit-def: $vgpr52
|
|
; VI-NEXT: ; implicit-def: $vgpr51
|
|
; VI-NEXT: ; implicit-def: $vgpr50
|
|
; VI-NEXT: ; implicit-def: $vgpr49
|
|
; VI-NEXT: ; implicit-def: $vgpr48
|
|
; VI-NEXT: ; implicit-def: $vgpr39
|
|
; VI-NEXT: ; implicit-def: $vgpr38
|
|
; VI-NEXT: ; implicit-def: $vgpr37
|
|
; VI-NEXT: ; implicit-def: $vgpr36
|
|
; VI-NEXT: ; implicit-def: $vgpr35
|
|
; VI-NEXT: ; implicit-def: $vgpr34
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: .LBB54_2: ; %Flow
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB54_4
|
|
; VI-NEXT: ; %bb.3: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; VI-NEXT: v_add_f16_sdwa v0, v59, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v59
|
|
; VI-NEXT: v_add_f16_sdwa v2, v58, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v58
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v2, v57, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v57
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v3, v56, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v56
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_add_f16_sdwa v4, v47, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v47
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_add_f16_sdwa v5, v46, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v46
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_add_f16_sdwa v6, v45, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v45
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_add_f16_sdwa v7, v44, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v44
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_add_f16_sdwa v8, v43, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v43
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_add_f16_sdwa v9, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, 0x200, v42
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v41, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, 0x200, v41
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_add_f16_sdwa v11, v40, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, 0x200, v40
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_add_f16_sdwa v12, v55, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, 0x200, v55
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_add_f16_sdwa v13, v54, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v54
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_add_f16_sdwa v14, v53, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v53
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v52, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v52
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v51, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v51
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v50, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v50
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v49, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v48, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v39, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v38, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v37, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v23, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v36, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: v_add_f16_sdwa v24, v35, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v25, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; VI-NEXT: v_add_f16_sdwa v25, v34, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v26, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; VI-NEXT: v_add_f16_sdwa v26, v33, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v26, v28, v26
|
|
; VI-NEXT: v_add_f16_sdwa v27, v32, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; VI-NEXT: .LBB54_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56f16_to_v14f64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_mov_b32_e32 v59, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v27
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v26
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v25
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v22
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v13
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v12
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v54
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v11
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v10
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v40
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v56, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v57, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v58, v1
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v41
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v47
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v56
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v57
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v58
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v59
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB54_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v37
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr47
|
|
; GFX9-NEXT: ; implicit-def: $vgpr46
|
|
; GFX9-NEXT: ; implicit-def: $vgpr45
|
|
; GFX9-NEXT: ; implicit-def: $vgpr44
|
|
; GFX9-NEXT: ; implicit-def: $vgpr43
|
|
; GFX9-NEXT: ; implicit-def: $vgpr42
|
|
; GFX9-NEXT: ; implicit-def: $vgpr41
|
|
; GFX9-NEXT: ; implicit-def: $vgpr40
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr54
|
|
; GFX9-NEXT: ; implicit-def: $vgpr53
|
|
; GFX9-NEXT: ; implicit-def: $vgpr52
|
|
; GFX9-NEXT: ; implicit-def: $vgpr51
|
|
; GFX9-NEXT: ; implicit-def: $vgpr50
|
|
; GFX9-NEXT: ; implicit-def: $vgpr49
|
|
; GFX9-NEXT: ; implicit-def: $vgpr48
|
|
; GFX9-NEXT: ; implicit-def: $vgpr39
|
|
; GFX9-NEXT: ; implicit-def: $vgpr38
|
|
; GFX9-NEXT: ; implicit-def: $vgpr37
|
|
; GFX9-NEXT: ; implicit-def: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr34
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr35
|
|
; GFX9-NEXT: ; implicit-def: $vgpr36
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; kill: killed $vgpr28
|
|
; GFX9-NEXT: .LBB54_2: ; %Flow
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB54_4
|
|
; GFX9-NEXT: ; %bb.3: ; %cmp.true
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v0, v36, v59, s6
|
|
; GFX9-NEXT: s_movk_i32 s7, 0x200
|
|
; GFX9-NEXT: v_perm_b32 v1, v35, v58, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v62, v57, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v34, v56, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v61, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v32, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v43, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v0, v0, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, v1, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, v2, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, v3, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, v4, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, v5, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v42, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v41, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v40, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v55, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v37, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, v12, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, v13, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, v14, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, v15, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, v16, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, v17, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, v18, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, v19, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, v20, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, v21, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, v22, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, v23, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v24, v25, v24, s6
|
|
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v24, v24, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v25, v26, v25, s6
|
|
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v25, v25, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v26, v27, v26, s6
|
|
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_pk_add_f16 v26, v26, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v27, v28, v27, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v27, v27, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB54_4: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v14f64:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB54_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: .LBB54_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56f16_to_v14f64:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v28
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v69, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v70, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v71, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v80, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v68, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v67, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v66, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v65, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v64, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v55, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v54, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v53, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v52, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v51, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v50, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v49, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v48, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v39, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v38, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v37, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v36, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v35, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v34, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v33, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v32, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v31, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v30, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v29, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB54_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: .LBB54_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <56 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <56 x half> %a1 to <14 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x half> %a to <14 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x double> %phi
|
|
}
|
|
|
|
define inreg <14 x double> @bitcast_v56f16_to_v14f64_scalar(<56 x half> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v56f16_to_v14f64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v32, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v32, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v32, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v32, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v32, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v32, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v32, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v32, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v32, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v32, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v32, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v32, s51, 11
|
|
; SI-NEXT: v_writelane_b32 v32, s52, 12
|
|
; SI-NEXT: v_writelane_b32 v32, s53, 13
|
|
; SI-NEXT: v_writelane_b32 v32, s54, 14
|
|
; SI-NEXT: v_writelane_b32 v32, s55, 15
|
|
; SI-NEXT: v_writelane_b32 v32, s64, 16
|
|
; SI-NEXT: v_writelane_b32 v32, s65, 17
|
|
; SI-NEXT: v_writelane_b32 v32, s66, 18
|
|
; SI-NEXT: v_writelane_b32 v32, s67, 19
|
|
; SI-NEXT: v_writelane_b32 v32, s68, 20
|
|
; SI-NEXT: v_writelane_b32 v32, s69, 21
|
|
; SI-NEXT: v_writelane_b32 v32, s70, 22
|
|
; SI-NEXT: v_writelane_b32 v32, s71, 23
|
|
; SI-NEXT: v_writelane_b32 v32, s80, 24
|
|
; SI-NEXT: v_writelane_b32 v32, s81, 25
|
|
; SI-NEXT: v_writelane_b32 v32, s82, 26
|
|
; SI-NEXT: v_writelane_b32 v32, s83, 27
|
|
; SI-NEXT: v_writelane_b32 v32, s84, 28
|
|
; SI-NEXT: v_writelane_b32 v32, s85, 29
|
|
; SI-NEXT: v_writelane_b32 v32, s86, 30
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s72, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s74, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s76, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s78, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s88, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s90, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s92, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s95, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s34, v0
|
|
; SI-NEXT: v_writelane_b32 v32, s87, 31
|
|
; SI-NEXT: s_lshr_b32 s94, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s69, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s70, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s71, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s80, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s81, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s82, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s83, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s84, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s85, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s86, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s87, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s7, s6, 16
|
|
; SI-NEXT: s_lshr_b32 s9, s8, 16
|
|
; SI-NEXT: s_lshr_b32 s11, s10, 16
|
|
; SI-NEXT: s_lshr_b32 s13, s12, 16
|
|
; SI-NEXT: s_lshr_b32 s15, s14, 16
|
|
; SI-NEXT: s_lshr_b32 s73, s72, 16
|
|
; SI-NEXT: s_lshr_b32 s75, s74, 16
|
|
; SI-NEXT: s_lshr_b32 s77, s76, 16
|
|
; SI-NEXT: s_lshr_b32 s79, s78, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s88, 16
|
|
; SI-NEXT: s_lshr_b32 s91, s90, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s92, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s95, 16
|
|
; SI-NEXT: s_lshr_b32 s68, s34, 16
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: s_cbranch_scc0 .LBB55_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; SI-NEXT: s_or_b32 s36, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; SI-NEXT: s_or_b32 s37, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; SI-NEXT: s_or_b32 s38, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; SI-NEXT: s_or_b32 s39, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s83, 16
|
|
; SI-NEXT: s_or_b32 s40, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; SI-NEXT: s_or_b32 s41, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s42, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; SI-NEXT: s_or_b32 s43, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s44, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s45, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s69, 16
|
|
; SI-NEXT: s_or_b32 s46, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s47, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; SI-NEXT: s_or_b32 s48, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s94, 16
|
|
; SI-NEXT: s_or_b32 s49, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s34, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s50, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s95, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; SI-NEXT: s_or_b32 s51, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s92, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_or_b32 s52, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s90, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s91, 16
|
|
; SI-NEXT: s_or_b32 s53, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s88, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s54, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s78, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; SI-NEXT: s_or_b32 s55, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s76, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s77, 16
|
|
; SI-NEXT: s_or_b32 s56, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s74, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; SI-NEXT: s_or_b32 s57, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s72, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s73, 16
|
|
; SI-NEXT: s_or_b32 s58, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; SI-NEXT: s_or_b32 s59, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; SI-NEXT: s_or_b32 s60, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s11, 16
|
|
; SI-NEXT: s_or_b32 s61, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s9, 16
|
|
; SI-NEXT: s_or_b32 s62, s4, s5
|
|
; SI-NEXT: s_and_b32 s4, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; SI-NEXT: s_or_b32 s63, s4, s5
|
|
; SI-NEXT: s_cbranch_execnz .LBB55_4
|
|
; SI-NEXT: .LBB55_2: ; %cmp.true
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, s87
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s86
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s17
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s85
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s18
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s84
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s19
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s83
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s20
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_or_b32_e32 v3, v5, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s82
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_or_b32_e32 v4, v6, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s81
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s22
|
|
; SI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s80
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s23
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_or_b32_e32 v6, v8, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s71
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v9, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s70
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s25
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s69
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s26
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_or_b32_e32 v9, v11, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v12, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s30
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s28
|
|
; SI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s94
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s29
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s68
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v13, v15, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s31
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s95
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s93
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s92
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s91
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v18, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s90
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s89
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s88
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s79
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s78
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s77
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v21, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s76
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s75
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s74
|
|
; SI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s73
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s72
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s13
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s12
|
|
; SI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s10
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: v_or_b32_e32 v24, v26, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v25, v27, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, s7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, s6
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v29, 0x38000000, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v27, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v27, v29, v27
|
|
; SI-NEXT: s_branch .LBB55_5
|
|
; SI-NEXT: .LBB55_3:
|
|
; SI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; SI-NEXT: s_branch .LBB55_2
|
|
; SI-NEXT: .LBB55_4:
|
|
; SI-NEXT: v_mov_b32_e32 v0, s36
|
|
; SI-NEXT: v_mov_b32_e32 v1, s37
|
|
; SI-NEXT: v_mov_b32_e32 v2, s38
|
|
; SI-NEXT: v_mov_b32_e32 v3, s39
|
|
; SI-NEXT: v_mov_b32_e32 v4, s40
|
|
; SI-NEXT: v_mov_b32_e32 v5, s41
|
|
; SI-NEXT: v_mov_b32_e32 v6, s42
|
|
; SI-NEXT: v_mov_b32_e32 v7, s43
|
|
; SI-NEXT: v_mov_b32_e32 v8, s44
|
|
; SI-NEXT: v_mov_b32_e32 v9, s45
|
|
; SI-NEXT: v_mov_b32_e32 v10, s46
|
|
; SI-NEXT: v_mov_b32_e32 v11, s47
|
|
; SI-NEXT: v_mov_b32_e32 v12, s48
|
|
; SI-NEXT: v_mov_b32_e32 v13, s49
|
|
; SI-NEXT: v_mov_b32_e32 v14, s50
|
|
; SI-NEXT: v_mov_b32_e32 v15, s51
|
|
; SI-NEXT: v_mov_b32_e32 v16, s52
|
|
; SI-NEXT: v_mov_b32_e32 v17, s53
|
|
; SI-NEXT: v_mov_b32_e32 v18, s54
|
|
; SI-NEXT: v_mov_b32_e32 v19, s55
|
|
; SI-NEXT: v_mov_b32_e32 v20, s56
|
|
; SI-NEXT: v_mov_b32_e32 v21, s57
|
|
; SI-NEXT: v_mov_b32_e32 v22, s58
|
|
; SI-NEXT: v_mov_b32_e32 v23, s59
|
|
; SI-NEXT: v_mov_b32_e32 v24, s60
|
|
; SI-NEXT: v_mov_b32_e32 v25, s61
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v27, s63
|
|
; SI-NEXT: v_mov_b32_e32 v28, s64
|
|
; SI-NEXT: v_mov_b32_e32 v29, s65
|
|
; SI-NEXT: v_mov_b32_e32 v30, s66
|
|
; SI-NEXT: v_mov_b32_e32 v31, s67
|
|
; SI-NEXT: .LBB55_5: ; %end
|
|
; SI-NEXT: v_readlane_b32 s87, v32, 31
|
|
; SI-NEXT: v_readlane_b32 s86, v32, 30
|
|
; SI-NEXT: v_readlane_b32 s85, v32, 29
|
|
; SI-NEXT: v_readlane_b32 s84, v32, 28
|
|
; SI-NEXT: v_readlane_b32 s83, v32, 27
|
|
; SI-NEXT: v_readlane_b32 s82, v32, 26
|
|
; SI-NEXT: v_readlane_b32 s81, v32, 25
|
|
; SI-NEXT: v_readlane_b32 s80, v32, 24
|
|
; SI-NEXT: v_readlane_b32 s71, v32, 23
|
|
; SI-NEXT: v_readlane_b32 s70, v32, 22
|
|
; SI-NEXT: v_readlane_b32 s69, v32, 21
|
|
; SI-NEXT: v_readlane_b32 s68, v32, 20
|
|
; SI-NEXT: v_readlane_b32 s67, v32, 19
|
|
; SI-NEXT: v_readlane_b32 s66, v32, 18
|
|
; SI-NEXT: v_readlane_b32 s65, v32, 17
|
|
; SI-NEXT: v_readlane_b32 s64, v32, 16
|
|
; SI-NEXT: v_readlane_b32 s55, v32, 15
|
|
; SI-NEXT: v_readlane_b32 s54, v32, 14
|
|
; SI-NEXT: v_readlane_b32 s53, v32, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v32, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v32, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v32, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v32, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v32, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v32, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v32, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v32, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v32, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v32, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v32, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v32, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v32, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56f16_to_v14f64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v32, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s31, 1
|
|
; VI-NEXT: v_writelane_b32 v32, s34, 2
|
|
; VI-NEXT: v_writelane_b32 v32, s35, 3
|
|
; VI-NEXT: v_writelane_b32 v32, s36, 4
|
|
; VI-NEXT: v_writelane_b32 v32, s37, 5
|
|
; VI-NEXT: v_writelane_b32 v32, s38, 6
|
|
; VI-NEXT: v_writelane_b32 v32, s39, 7
|
|
; VI-NEXT: v_writelane_b32 v32, s48, 8
|
|
; VI-NEXT: v_writelane_b32 v32, s49, 9
|
|
; VI-NEXT: v_writelane_b32 v32, s50, 10
|
|
; VI-NEXT: v_writelane_b32 v32, s51, 11
|
|
; VI-NEXT: v_writelane_b32 v32, s52, 12
|
|
; VI-NEXT: v_writelane_b32 v32, s53, 13
|
|
; VI-NEXT: v_writelane_b32 v32, s54, 14
|
|
; VI-NEXT: v_writelane_b32 v32, s55, 15
|
|
; VI-NEXT: v_writelane_b32 v32, s64, 16
|
|
; VI-NEXT: v_writelane_b32 v32, s65, 17
|
|
; VI-NEXT: v_writelane_b32 v32, s66, 18
|
|
; VI-NEXT: v_writelane_b32 v32, s67, 19
|
|
; VI-NEXT: v_writelane_b32 v32, s68, 20
|
|
; VI-NEXT: v_writelane_b32 v32, s69, 21
|
|
; VI-NEXT: v_writelane_b32 v32, s70, 22
|
|
; VI-NEXT: v_writelane_b32 v32, s71, 23
|
|
; VI-NEXT: v_writelane_b32 v32, s80, 24
|
|
; VI-NEXT: v_writelane_b32 v32, s81, 25
|
|
; VI-NEXT: v_writelane_b32 v32, s82, 26
|
|
; VI-NEXT: v_writelane_b32 v32, s83, 27
|
|
; VI-NEXT: v_writelane_b32 v32, s84, 28
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; VI-NEXT: v_writelane_b32 v32, s85, 29
|
|
; VI-NEXT: s_lshr_b32 s15, s8, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v11
|
|
; VI-NEXT: ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
|
|
; VI-NEXT: v_writelane_b32 v32, s86, 30
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v13
|
|
; VI-NEXT: s_lshr_b32 s61, s10, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s72, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s74, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s76, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s79, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s91, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s34, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s69, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s80, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s83, v0
|
|
; VI-NEXT: v_writelane_b32 v33, s15, 0
|
|
; VI-NEXT: v_writelane_b32 v32, s87, 31
|
|
; VI-NEXT: s_lshr_b32 s56, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s90, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s68, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s70, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s81, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s84, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s86, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s87, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s7, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s9, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s11, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s13, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s72, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s74, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s76, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s79, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s91, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s34, 16
|
|
; VI-NEXT: s_lshr_b32 s71, s69, 16
|
|
; VI-NEXT: s_lshr_b32 s82, s80, 16
|
|
; VI-NEXT: s_lshr_b32 s85, s83, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: v_writelane_b32 v33, s61, 1
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: v_writelane_b32 v33, s60, 2
|
|
; VI-NEXT: v_writelane_b32 v33, s59, 3
|
|
; VI-NEXT: s_cbranch_scc0 .LBB55_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s13, 16
|
|
; VI-NEXT: s_or_b32 s36, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s5, s9, 16
|
|
; VI-NEXT: s_or_b32 s38, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s5, s7, 16
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s11, 16
|
|
; VI-NEXT: s_or_b32 s39, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s5, s87, 16
|
|
; VI-NEXT: s_or_b32 s37, s40, s41
|
|
; VI-NEXT: s_or_b32 s40, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s5, s86, 16
|
|
; VI-NEXT: s_or_b32 s41, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s5, s84, 16
|
|
; VI-NEXT: s_or_b32 s42, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; VI-NEXT: s_or_b32 s43, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; VI-NEXT: s_or_b32 s44, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; VI-NEXT: s_or_b32 s45, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; VI-NEXT: s_or_b32 s46, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s5, s90, 16
|
|
; VI-NEXT: s_or_b32 s47, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s5, s75, 16
|
|
; VI-NEXT: s_or_b32 s48, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s5, s56, 16
|
|
; VI-NEXT: s_or_b32 s49, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s83
|
|
; VI-NEXT: s_lshl_b32 s5, s85, 16
|
|
; VI-NEXT: s_or_b32 s50, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s80
|
|
; VI-NEXT: s_lshl_b32 s5, s82, 16
|
|
; VI-NEXT: s_or_b32 s51, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s69
|
|
; VI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; VI-NEXT: s_or_b32 s52, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s34
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s53, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s91
|
|
; VI-NEXT: s_lshl_b32 s5, s30, 16
|
|
; VI-NEXT: s_or_b32 s54, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s79
|
|
; VI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; VI-NEXT: s_or_b32 s55, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s76
|
|
; VI-NEXT: s_lshl_b32 s5, s62, 16
|
|
; VI-NEXT: s_mov_b32 s77, s56
|
|
; VI-NEXT: s_or_b32 s56, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s74
|
|
; VI-NEXT: s_lshl_b32 s5, s57, 16
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: s_mov_b32 s75, s57
|
|
; VI-NEXT: s_or_b32 s57, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s72
|
|
; VI-NEXT: s_lshl_b32 s5, s58, 16
|
|
; VI-NEXT: s_mov_b32 s78, s58
|
|
; VI-NEXT: s_or_b32 s58, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s5, s59, 16
|
|
; VI-NEXT: s_or_b32 s59, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s5, s60, 16
|
|
; VI-NEXT: s_or_b32 s60, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s5, s61, 16
|
|
; VI-NEXT: s_or_b32 s61, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s5, s15, 16
|
|
; VI-NEXT: s_mov_b32 s88, s62
|
|
; VI-NEXT: s_or_b32 s62, s4, s5
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s5, s63, 16
|
|
; VI-NEXT: s_mov_b32 s15, s63
|
|
; VI-NEXT: s_or_b32 s63, s4, s5
|
|
; VI-NEXT: s_cbranch_execnz .LBB55_4
|
|
; VI-NEXT: .LBB55_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; VI-NEXT: v_mov_b32_e32 v0, s13
|
|
; VI-NEXT: v_mov_b32_e32 v2, s11
|
|
; VI-NEXT: v_add_f16_sdwa v0, v0, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, s16, v27
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s17, v27
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_or_b32_e32 v1, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v2, s9
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s18, v27
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v3, s7
|
|
; VI-NEXT: v_add_f16_sdwa v3, v3, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, s19, v27
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_mov_b32_e32 v4, s87
|
|
; VI-NEXT: v_add_f16_sdwa v4, v4, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, s20, v27
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_mov_b32_e32 v5, s86
|
|
; VI-NEXT: v_add_f16_sdwa v5, v5, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, s21, v27
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_mov_b32_e32 v6, s84
|
|
; VI-NEXT: v_add_f16_sdwa v6, v6, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, s22, v27
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_mov_b32_e32 v7, s81
|
|
; VI-NEXT: v_add_f16_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, s23, v27
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_mov_b32_e32 v8, s70
|
|
; VI-NEXT: v_add_f16_sdwa v8, v8, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, s24, v27
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_mov_b32_e32 v9, s68
|
|
; VI-NEXT: v_add_f16_sdwa v9, v9, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, s25, v27
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_mov_b32_e32 v10, s31
|
|
; VI-NEXT: v_add_f16_sdwa v10, v10, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, s26, v27
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_mov_b32_e32 v11, s90
|
|
; VI-NEXT: v_add_f16_sdwa v11, v11, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, s27, v27
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_mov_b32_e32 v12, s73
|
|
; VI-NEXT: v_add_f16_sdwa v12, v12, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, s28, v27
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_mov_b32_e32 v13, s77
|
|
; VI-NEXT: v_add_f16_sdwa v13, v13, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, s29, v27
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_mov_b32_e32 v14, s85
|
|
; VI-NEXT: v_add_f16_sdwa v14, v14, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, s83, v27
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_mov_b32_e32 v15, s82
|
|
; VI-NEXT: v_add_f16_sdwa v15, v15, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, s80, v27
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_mov_b32_e32 v16, s71
|
|
; VI-NEXT: v_add_f16_sdwa v16, v16, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, s69, v27
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_mov_b32_e32 v17, s35
|
|
; VI-NEXT: v_add_f16_sdwa v17, v17, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, s34, v27
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_mov_b32_e32 v18, s30
|
|
; VI-NEXT: v_add_f16_sdwa v18, v18, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, s91, v27
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_mov_b32_e32 v19, s89
|
|
; VI-NEXT: v_add_f16_sdwa v19, v19, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, s79, v27
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_mov_b32_e32 v20, s88
|
|
; VI-NEXT: v_add_f16_sdwa v20, v20, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, s76, v27
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_mov_b32_e32 v21, s75
|
|
; VI-NEXT: v_add_f16_sdwa v21, v21, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, s74, v27
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_mov_b32_e32 v22, s78
|
|
; VI-NEXT: v_add_f16_sdwa v22, v22, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v23, s72, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 3
|
|
; VI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; VI-NEXT: v_mov_b32_e32 v23, s4
|
|
; VI-NEXT: v_add_f16_sdwa v23, v23, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, s14, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 2
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: v_mov_b32_e32 v24, s4
|
|
; VI-NEXT: v_add_f16_sdwa v24, v24, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v25, s12, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 1
|
|
; VI-NEXT: v_or_b32_e32 v24, v25, v24
|
|
; VI-NEXT: v_mov_b32_e32 v25, s4
|
|
; VI-NEXT: v_add_f16_sdwa v25, v25, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v26, s10, v27
|
|
; VI-NEXT: v_readlane_b32 s4, v33, 0
|
|
; VI-NEXT: v_or_b32_e32 v25, v26, v25
|
|
; VI-NEXT: v_mov_b32_e32 v26, s4
|
|
; VI-NEXT: v_add_f16_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v28, s8, v27
|
|
; VI-NEXT: v_or_b32_e32 v26, v28, v26
|
|
; VI-NEXT: v_mov_b32_e32 v28, s15
|
|
; VI-NEXT: v_add_f16_sdwa v28, v28, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v27, s6, v27
|
|
; VI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; VI-NEXT: s_branch .LBB55_5
|
|
; VI-NEXT: .LBB55_3:
|
|
; VI-NEXT: s_mov_b32 s73, s75
|
|
; VI-NEXT: s_mov_b32 s88, s62
|
|
; VI-NEXT: s_mov_b32 s77, s56
|
|
; VI-NEXT: s_mov_b32 s75, s57
|
|
; VI-NEXT: s_mov_b32 s78, s58
|
|
; VI-NEXT: s_mov_b32 s15, s63
|
|
; VI-NEXT: ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
|
|
; VI-NEXT: s_branch .LBB55_2
|
|
; VI-NEXT: .LBB55_4:
|
|
; VI-NEXT: v_mov_b32_e32 v0, s36
|
|
; VI-NEXT: v_mov_b32_e32 v1, s37
|
|
; VI-NEXT: v_mov_b32_e32 v2, s38
|
|
; VI-NEXT: v_mov_b32_e32 v3, s39
|
|
; VI-NEXT: v_mov_b32_e32 v4, s40
|
|
; VI-NEXT: v_mov_b32_e32 v5, s41
|
|
; VI-NEXT: v_mov_b32_e32 v6, s42
|
|
; VI-NEXT: v_mov_b32_e32 v7, s43
|
|
; VI-NEXT: v_mov_b32_e32 v8, s44
|
|
; VI-NEXT: v_mov_b32_e32 v9, s45
|
|
; VI-NEXT: v_mov_b32_e32 v10, s46
|
|
; VI-NEXT: v_mov_b32_e32 v11, s47
|
|
; VI-NEXT: v_mov_b32_e32 v12, s48
|
|
; VI-NEXT: v_mov_b32_e32 v13, s49
|
|
; VI-NEXT: v_mov_b32_e32 v14, s50
|
|
; VI-NEXT: v_mov_b32_e32 v15, s51
|
|
; VI-NEXT: v_mov_b32_e32 v16, s52
|
|
; VI-NEXT: v_mov_b32_e32 v17, s53
|
|
; VI-NEXT: v_mov_b32_e32 v18, s54
|
|
; VI-NEXT: v_mov_b32_e32 v19, s55
|
|
; VI-NEXT: v_mov_b32_e32 v20, s56
|
|
; VI-NEXT: v_mov_b32_e32 v21, s57
|
|
; VI-NEXT: v_mov_b32_e32 v22, s58
|
|
; VI-NEXT: v_mov_b32_e32 v23, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s60
|
|
; VI-NEXT: v_mov_b32_e32 v25, s61
|
|
; VI-NEXT: v_mov_b32_e32 v26, s62
|
|
; VI-NEXT: v_mov_b32_e32 v27, s63
|
|
; VI-NEXT: v_mov_b32_e32 v28, s64
|
|
; VI-NEXT: v_mov_b32_e32 v29, s65
|
|
; VI-NEXT: v_mov_b32_e32 v30, s66
|
|
; VI-NEXT: v_mov_b32_e32 v31, s67
|
|
; VI-NEXT: .LBB55_5: ; %end
|
|
; VI-NEXT: v_readlane_b32 s87, v32, 31
|
|
; VI-NEXT: v_readlane_b32 s86, v32, 30
|
|
; VI-NEXT: v_readlane_b32 s85, v32, 29
|
|
; VI-NEXT: v_readlane_b32 s84, v32, 28
|
|
; VI-NEXT: v_readlane_b32 s83, v32, 27
|
|
; VI-NEXT: v_readlane_b32 s82, v32, 26
|
|
; VI-NEXT: v_readlane_b32 s81, v32, 25
|
|
; VI-NEXT: v_readlane_b32 s80, v32, 24
|
|
; VI-NEXT: v_readlane_b32 s71, v32, 23
|
|
; VI-NEXT: v_readlane_b32 s70, v32, 22
|
|
; VI-NEXT: v_readlane_b32 s69, v32, 21
|
|
; VI-NEXT: v_readlane_b32 s68, v32, 20
|
|
; VI-NEXT: v_readlane_b32 s67, v32, 19
|
|
; VI-NEXT: v_readlane_b32 s66, v32, 18
|
|
; VI-NEXT: v_readlane_b32 s65, v32, 17
|
|
; VI-NEXT: v_readlane_b32 s64, v32, 16
|
|
; VI-NEXT: v_readlane_b32 s55, v32, 15
|
|
; VI-NEXT: v_readlane_b32 s54, v32, 14
|
|
; VI-NEXT: v_readlane_b32 s53, v32, 13
|
|
; VI-NEXT: v_readlane_b32 s52, v32, 12
|
|
; VI-NEXT: v_readlane_b32 s51, v32, 11
|
|
; VI-NEXT: v_readlane_b32 s50, v32, 10
|
|
; VI-NEXT: v_readlane_b32 s49, v32, 9
|
|
; VI-NEXT: v_readlane_b32 s48, v32, 8
|
|
; VI-NEXT: v_readlane_b32 s39, v32, 7
|
|
; VI-NEXT: v_readlane_b32 s38, v32, 6
|
|
; VI-NEXT: v_readlane_b32 s37, v32, 5
|
|
; VI-NEXT: v_readlane_b32 s36, v32, 4
|
|
; VI-NEXT: v_readlane_b32 s35, v32, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v32, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v32, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v32, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56f16_to_v14f64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: v_writelane_b32 v32, s36, 0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s37, 1
|
|
; GFX9-NEXT: v_writelane_b32 v32, s38, 2
|
|
; GFX9-NEXT: v_writelane_b32 v32, s39, 3
|
|
; GFX9-NEXT: v_writelane_b32 v32, s48, 4
|
|
; GFX9-NEXT: v_writelane_b32 v32, s49, 5
|
|
; GFX9-NEXT: v_writelane_b32 v32, s50, 6
|
|
; GFX9-NEXT: v_writelane_b32 v32, s51, 7
|
|
; GFX9-NEXT: v_writelane_b32 v32, s52, 8
|
|
; GFX9-NEXT: s_lshr_b32 s12, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s13, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s14, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s15, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX9-NEXT: v_writelane_b32 v32, s53, 9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s63, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s62, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s61, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s60, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s59, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s58, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s57, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s56, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s88, v5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s36, s16, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s37, s17, s40
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s38, s18, s15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v3
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s39, s19, s14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v2
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s20, s13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v1
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s41, s21, s12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v0
|
|
; GFX9-NEXT: v_writelane_b32 v32, s54, 10
|
|
; GFX9-NEXT: s_lshr_b32 s4, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s5, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s6, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s7, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s8, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s9, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s10, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s63, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s62, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s61, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s60, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s59, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s58, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s57, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s56, 16
|
|
; GFX9-NEXT: s_lshr_b32 s89, s88, 16
|
|
; GFX9-NEXT: s_lshr_b32 s17, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s18, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s19, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s20, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s21, s12, 16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s42, v14
|
|
; GFX9-NEXT: v_writelane_b32 v32, s55, 11
|
|
; GFX9-NEXT: s_cmp_lg_u32 s42, 0
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s42, s22, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s23, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s44, s24, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s45, s25, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s46, s26, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s47, s27, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s48, s28, s5
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s49, s29, s4
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s50, s12, s21
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s51, s13, s20
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s52, s14, s19
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s53, s15, s18
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s54, s16, s17
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s55, s88, s89
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s56, s56, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s57, s57, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s58, s58, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s59, s59, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s60, s60, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s61, s61, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s62, s62, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s63, s63, s72
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB55_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB55_4
|
|
; GFX9-NEXT: .LBB55_2: ; %cmp.true
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, 0x200
|
|
; GFX9-NEXT: v_pk_add_f16 v0, s36, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, s37, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, s38, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, s39, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, s40, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, s41, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, s42, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, s43, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, s44, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v9, s45, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, s46, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, s47, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, s48, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, s49, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, s50, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, s51, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, s52, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, s53, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, s54, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, s55, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, s56, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, s57, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, s58, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, s59, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v24, s60, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v25, s61, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v26, s62, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v27, s63, v27 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_branch .LBB55_5
|
|
; GFX9-NEXT: .LBB55_3:
|
|
; GFX9-NEXT: s_branch .LBB55_2
|
|
; GFX9-NEXT: .LBB55_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s36
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s37
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s38
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s39
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s48
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s49
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s50
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s51
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s52
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s53
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s54
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s55
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s64
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s65
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s66
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s67
|
|
; GFX9-NEXT: .LBB55_5: ; %end
|
|
; GFX9-NEXT: v_readlane_b32 s55, v32, 11
|
|
; GFX9-NEXT: v_readlane_b32 s54, v32, 10
|
|
; GFX9-NEXT: v_readlane_b32 s53, v32, 9
|
|
; GFX9-NEXT: v_readlane_b32 s52, v32, 8
|
|
; GFX9-NEXT: v_readlane_b32 s51, v32, 7
|
|
; GFX9-NEXT: v_readlane_b32 s50, v32, 6
|
|
; GFX9-NEXT: v_readlane_b32 s49, v32, 5
|
|
; GFX9-NEXT: v_readlane_b32 s48, v32, 4
|
|
; GFX9-NEXT: v_readlane_b32 s39, v32, 3
|
|
; GFX9-NEXT: v_readlane_b32 s38, v32, 2
|
|
; GFX9-NEXT: v_readlane_b32 s37, v32, 1
|
|
; GFX9-NEXT: v_readlane_b32 s36, v32, 0
|
|
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v56f16_to_v14f64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_readfirstlane_b32 s44, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s56, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s58, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s59, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s60, v5
|
|
; GFX11-NEXT: v_readfirstlane_b32 s62, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s74, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s76, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s77, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s78, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s88, v10
|
|
; GFX11-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s0, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s44, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s56, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s58, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s59, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s60, 16
|
|
; GFX11-NEXT: s_lshr_b32 s75, s62, 16
|
|
; GFX11-NEXT: s_lshr_b32 s79, s74, 16
|
|
; GFX11-NEXT: s_lshr_b32 s89, s76, 16
|
|
; GFX11-NEXT: s_lshr_b32 s90, s77, 16
|
|
; GFX11-NEXT: s_lshr_b32 s91, s78, 16
|
|
; GFX11-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-NEXT: s_cmp_lg_u32 s88, 0
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s0, s0, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s78, s91
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s77, s90
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s76, s89
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s74, s79
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s62, s75
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s60, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s24, s59, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s25, s58, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s26, s56, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s27, s44, s57
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB55_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB55_4
|
|
; GFX11-NEXT: .LBB55_2: ; %cmp.true
|
|
; GFX11-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v18, 0x200, s18 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v19, 0x200, s19 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v20, 0x200, s20 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v21, 0x200, s21 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v22, 0x200, s22 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v23, 0x200, s23 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v24, 0x200, s24 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v25, 0x200, s25 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v26, 0x200, s26 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: v_pk_add_f16 v27, 0x200, s27 op_sel_hi:[0,1]
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB55_3:
|
|
; GFX11-NEXT: s_branch .LBB55_2
|
|
; GFX11-NEXT: .LBB55_4:
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
|
|
; GFX11-NEXT: v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <56 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <56 x half> %a1 to <14 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x half> %a to <14 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <14 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <14 x double> %phi
|
|
}
|
|
|
|
define <56 x half> @bitcast_v56i16_to_v56f16(<56 x i16> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v56i16_to_v56f16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v0
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v54
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v3
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v53
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v5
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v51
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v4
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v49
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v7
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v50
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v6
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v39
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v9
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v48
|
|
; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v11
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v60
|
|
; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v13
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v58
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v57
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v15
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v26
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v2
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: v_lshlrev_b32_e32 v43, 16, v52
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v34
|
|
; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v33
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v47
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v45
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v44
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v40
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v55
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; kill: killed $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB56_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v54
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_or_b32_e32 v60, v1, v28
|
|
; SI-NEXT: v_mov_b32_e32 v28, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v43
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_alignbit_b32 v1, v60, v43, 16
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v58, v1, v3
|
|
; SI-NEXT: v_alignbit_b32 v1, v58, v39, 16
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v5
|
|
; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v57, v1, v3
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v43
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v2
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v39
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v4
|
|
; SI-NEXT: v_mov_b32_e32 v39, v28
|
|
; SI-NEXT: ; implicit-def: $vgpr2
|
|
; SI-NEXT: ; implicit-def: $vgpr4
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_alignbit_b32 v1, v57, v5, 16
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v7
|
|
; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_or_b32_e32 v47, v1, v3
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v5
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v6
|
|
; SI-NEXT: ; implicit-def: $vgpr5
|
|
; SI-NEXT: ; implicit-def: $vgpr6
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_alignbit_b32 v1, v47, v7, 16
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_or_b32_e32 v45, v1, v3
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_alignbit_b32 v1, v45, v46, 16
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v11
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v7
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v8
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v46
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v10
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v56
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v12
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v59
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v14
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v62
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v16
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v29
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v18
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v30
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v31
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v32
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v63
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v61
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr0
|
|
; SI-NEXT: ; implicit-def: $vgpr7
|
|
; SI-NEXT: ; implicit-def: $vgpr8
|
|
; SI-NEXT: ; implicit-def: $vgpr9
|
|
; SI-NEXT: ; implicit-def: $vgpr10
|
|
; SI-NEXT: ; implicit-def: $vgpr11
|
|
; SI-NEXT: ; implicit-def: $vgpr12
|
|
; SI-NEXT: ; implicit-def: $vgpr14
|
|
; SI-NEXT: ; implicit-def: $vgpr16
|
|
; SI-NEXT: ; implicit-def: $vgpr18
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: ; implicit-def: $vgpr22
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_or_b32_e32 v44, v1, v3
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_alignbit_b32 v1, v44, v56, 16
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v13
|
|
; SI-NEXT: ; implicit-def: $vgpr13
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_or_b32_e32 v42, v1, v3
|
|
; SI-NEXT: v_alignbit_b32 v1, v42, v59, 16
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v15
|
|
; SI-NEXT: v_or_b32_e32 v40, v1, v35
|
|
; SI-NEXT: v_alignbit_b32 v1, v40, v62, 16
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v17
|
|
; SI-NEXT: v_or_b32_e32 v55, v1, v36
|
|
; SI-NEXT: v_alignbit_b32 v1, v55, v29, 16
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v19
|
|
; SI-NEXT: v_or_b32_e32 v53, v1, v34
|
|
; SI-NEXT: v_alignbit_b32 v1, v53, v30, 16
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v21
|
|
; SI-NEXT: v_or_b32_e32 v52, v1, v41
|
|
; SI-NEXT: v_alignbit_b32 v1, v52, v31, 16
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v23
|
|
; SI-NEXT: ; kill: killed $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: v_or_b32_e32 v50, v1, v37
|
|
; SI-NEXT: ; kill: killed $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: v_alignbit_b32 v1, v50, v32, 16
|
|
; SI-NEXT: ; kill: killed $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v25
|
|
; SI-NEXT: ; kill: killed $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: v_or_b32_e32 v48, v1, v33
|
|
; SI-NEXT: ; kill: killed $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: v_alignbit_b32 v1, v48, v63, 16
|
|
; SI-NEXT: ; kill: killed $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v27
|
|
; SI-NEXT: ; kill: killed $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: v_or_b32_e32 v3, v1, v38
|
|
; SI-NEXT: ; kill: killed $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: v_alignbit_b32 v1, v3, v61, 16
|
|
; SI-NEXT: ; kill: killed $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_mov_b32_e32 v28, v3
|
|
; SI-NEXT: ; implicit-def: $vgpr1
|
|
; SI-NEXT: ; implicit-def: $vgpr3
|
|
; SI-NEXT: ; implicit-def: $vgpr15
|
|
; SI-NEXT: ; implicit-def: $vgpr17
|
|
; SI-NEXT: ; implicit-def: $vgpr19
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr23
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; kill: killed $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: .LBB56_2: ; %Flow
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB56_4
|
|
; SI-NEXT: ; %bb.3: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v61, v26
|
|
; SI-NEXT: v_add_i32_e32 v39, vcc, 0x30000, v26
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v27
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: s_mov_b32 s6, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v26, v38, v26
|
|
; SI-NEXT: v_or_b32_e32 v24, v63, v24
|
|
; SI-NEXT: v_add_i32_e32 v28, vcc, s6, v26
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, s6, v24
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v25
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v24, v33, v24
|
|
; SI-NEXT: v_or_b32_e32 v22, v32, v22
|
|
; SI-NEXT: v_add_i32_e32 v48, vcc, s6, v24
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, s6, v22
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v23
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v22, v37, v22
|
|
; SI-NEXT: v_or_b32_e32 v20, v31, v20
|
|
; SI-NEXT: v_add_i32_e32 v50, vcc, s6, v22
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v20
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_or_b32_e32 v20, v41, v20
|
|
; SI-NEXT: v_or_b32_e32 v18, v30, v18
|
|
; SI-NEXT: v_add_i32_e32 v52, vcc, s6, v20
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, s6, v18
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_or_b32_e32 v18, v34, v18
|
|
; SI-NEXT: v_or_b32_e32 v16, v29, v16
|
|
; SI-NEXT: v_add_i32_e32 v53, vcc, s6, v18
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, s6, v16
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_or_b32_e32 v16, v36, v16
|
|
; SI-NEXT: v_or_b32_e32 v14, v62, v14
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v55, vcc, s6, v16
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, s6, v14
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_or_b32_e32 v14, v35, v14
|
|
; SI-NEXT: v_or_b32_e32 v12, v59, v12
|
|
; SI-NEXT: v_add_i32_e32 v40, vcc, s6, v14
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, s6, v12
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v13
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v56, v10
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v46, v8
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v43, v2
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v0
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; SI-NEXT: v_add_i32_e32 v42, vcc, s6, v12
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, s6, v10
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v11
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_add_i32_e32 v44, vcc, s6, v10
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, s6, v8
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v9
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: v_add_i32_e32 v45, vcc, s6, v8
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v6, v8, v6
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v6
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v7
|
|
; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; SI-NEXT: v_add_i32_e32 v47, vcc, s6, v6
|
|
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v4, v6, v4
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, s6, v4
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v5
|
|
; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; SI-NEXT: v_add_i32_e32 v57, vcc, s6, v4
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, s6, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v3
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v57
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: v_add_i32_e32 v58, vcc, s6, v2
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v58
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, s6, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v1
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; SI-NEXT: v_add_i32_e32 v60, vcc, s6, v0
|
|
; SI-NEXT: v_alignbit_b32 v0, v60, v2, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v58, v4, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v57, v6, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v47, v8, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v45, v10, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v44, v12, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v42, v14, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v40, v16, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v55, v18, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v53, v20, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v52, v22, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v50, v24, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v48, v26, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_alignbit_b32 v0, v28, v39, 16
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v45
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v44
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v42
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v40
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v55
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v53
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v52
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v50
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v60
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v47
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v28
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: .LBB56_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v49
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v39
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v60
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v58
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v57
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v47
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v45
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v11
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v44
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v13
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v42
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v15
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v40
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v17
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v55
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v19
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v53
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v21
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v52
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v23
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v50
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v48
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v27
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v28
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56i16_to_v56f16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB56_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_u16_e32 v0, 3, v0
|
|
; VI-NEXT: v_add_u16_e32 v56, 3, v56
|
|
; VI-NEXT: v_add_u16_e32 v1, 3, v1
|
|
; VI-NEXT: v_add_u16_e32 v47, 3, v47
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v2
|
|
; VI-NEXT: v_add_u16_e32 v46, 3, v46
|
|
; VI-NEXT: v_add_u16_e32 v3, 3, v3
|
|
; VI-NEXT: v_add_u16_e32 v45, 3, v45
|
|
; VI-NEXT: v_add_u16_e32 v4, 3, v4
|
|
; VI-NEXT: v_add_u16_e32 v44, 3, v44
|
|
; VI-NEXT: v_add_u16_e32 v5, 3, v5
|
|
; VI-NEXT: v_add_u16_e32 v43, 3, v43
|
|
; VI-NEXT: v_add_u16_e32 v6, 3, v6
|
|
; VI-NEXT: v_add_u16_e32 v42, 3, v42
|
|
; VI-NEXT: v_add_u16_e32 v7, 3, v7
|
|
; VI-NEXT: v_add_u16_e32 v41, 3, v41
|
|
; VI-NEXT: v_add_u16_e32 v8, 3, v8
|
|
; VI-NEXT: v_add_u16_e32 v40, 3, v40
|
|
; VI-NEXT: v_add_u16_e32 v9, 3, v9
|
|
; VI-NEXT: v_add_u16_e32 v55, 3, v55
|
|
; VI-NEXT: v_add_u16_e32 v10, 3, v10
|
|
; VI-NEXT: v_add_u16_e32 v54, 3, v54
|
|
; VI-NEXT: v_add_u16_e32 v11, 3, v11
|
|
; VI-NEXT: v_add_u16_e32 v53, 3, v53
|
|
; VI-NEXT: v_add_u16_e32 v12, 3, v12
|
|
; VI-NEXT: v_add_u16_e32 v52, 3, v52
|
|
; VI-NEXT: v_add_u16_e32 v13, 3, v13
|
|
; VI-NEXT: v_add_u16_e32 v51, 3, v51
|
|
; VI-NEXT: v_add_u16_e32 v14, 3, v14
|
|
; VI-NEXT: v_add_u16_e32 v50, 3, v50
|
|
; VI-NEXT: v_add_u16_e32 v15, 3, v15
|
|
; VI-NEXT: v_add_u16_e32 v49, 3, v49
|
|
; VI-NEXT: v_add_u16_e32 v16, 3, v16
|
|
; VI-NEXT: v_add_u16_e32 v48, 3, v48
|
|
; VI-NEXT: v_add_u16_e32 v17, 3, v17
|
|
; VI-NEXT: v_add_u16_e32 v39, 3, v39
|
|
; VI-NEXT: v_add_u16_e32 v18, 3, v18
|
|
; VI-NEXT: v_add_u16_e32 v38, 3, v38
|
|
; VI-NEXT: v_add_u16_e32 v19, 3, v19
|
|
; VI-NEXT: v_add_u16_e32 v37, 3, v37
|
|
; VI-NEXT: v_add_u16_e32 v20, 3, v20
|
|
; VI-NEXT: v_add_u16_e32 v36, 3, v36
|
|
; VI-NEXT: v_add_u16_e32 v21, 3, v21
|
|
; VI-NEXT: v_add_u16_e32 v35, 3, v35
|
|
; VI-NEXT: v_add_u16_e32 v22, 3, v22
|
|
; VI-NEXT: v_add_u16_e32 v34, 3, v34
|
|
; VI-NEXT: v_add_u16_e32 v23, 3, v23
|
|
; VI-NEXT: v_add_u16_e32 v33, 3, v33
|
|
; VI-NEXT: v_add_u16_e32 v24, 3, v24
|
|
; VI-NEXT: v_add_u16_e32 v32, 3, v32
|
|
; VI-NEXT: v_add_u16_e32 v25, 3, v25
|
|
; VI-NEXT: v_add_u16_e32 v31, 3, v31
|
|
; VI-NEXT: v_add_u16_e32 v26, 3, v26
|
|
; VI-NEXT: v_add_u16_e32 v30, 3, v30
|
|
; VI-NEXT: v_add_u16_e32 v27, 3, v27
|
|
; VI-NEXT: v_add_u16_e32 v29, 3, v29
|
|
; VI-NEXT: .LBB56_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v56
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v47
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v46
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v45
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v44
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v43
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v42
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v41
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v40
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v55
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v54
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v53
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v52
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v51
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v50
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v49
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v48
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v39
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v38
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v37
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v36
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v35
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v34
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v33
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v32
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v31
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v30
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v29
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56i16_to_v56f16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v0
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB56_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v27, v56, v27, s6
|
|
; GFX9-NEXT: v_perm_b32 v26, v47, v26, s6
|
|
; GFX9-NEXT: v_perm_b32 v25, v46, v25, s6
|
|
; GFX9-NEXT: v_perm_b32 v24, v45, v24, s6
|
|
; GFX9-NEXT: v_perm_b32 v23, v44, v23, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v43, v22, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v42, v21, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v41, v20, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v40, v19, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v55, v18, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v54, v17, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v53, v16, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v52, v15, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v51, v14, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v49, v12, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v48, v11, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v39, v10, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v37, v8, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v36, v7, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v35, v6, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v34, v5, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v4, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v32, v3, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v31, v2, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v30, v1, s6
|
|
; GFX9-NEXT: v_perm_b32 v0, v29, v0, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v27
|
|
; GFX9-NEXT: .LBB56_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v19, v40, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v41, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v42, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v43, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v44, v23, s4
|
|
; GFX9-NEXT: v_perm_b32 v24, v45, v24, s4
|
|
; GFX9-NEXT: v_perm_b32 v25, v46, v25, s4
|
|
; GFX9-NEXT: v_perm_b32 v26, v47, v26, s4
|
|
; GFX9-NEXT: v_perm_b32 v27, v56, v27, s4
|
|
; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v0, v29, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v30, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v31, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v32, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v34, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v35, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v36, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v37, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v39, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v48, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v49, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v51, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v52, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v53, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v54, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v55, v18, s4
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v56f16:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB56_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: .LBB56_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56i16_to_v56f16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v0
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-FAKE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB56_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v80, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v71, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v70, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v69, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v68, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v67, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v66, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v65, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v64, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v55, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v54, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v53, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v52, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v51, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v49, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v48, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v39, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v37, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v36, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v35, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v29, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v30, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v31, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v32, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v33, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v27
|
|
; GFX11-FAKE16-NEXT: .LBB56_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v29, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v30, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v31, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v32, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v33, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v35, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v36, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v37, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v39, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v48, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v49, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v51, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v52, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v53, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v54, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v55, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v64, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v65, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v66, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v67, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v68, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v69, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v70, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v71, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v80, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <56 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <56 x i16> %a1 to <56 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x i16> %a to <56 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x half> %phi
|
|
}
|
|
|
|
define inreg <56 x half> @bitcast_v56i16_to_v56f16_scalar(<56 x i16> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v56i16_to_v56f16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; SI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; SI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; SI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; SI-NEXT: v_writelane_b32 v28, s36, 4
|
|
; SI-NEXT: v_writelane_b32 v28, s37, 5
|
|
; SI-NEXT: v_writelane_b32 v28, s38, 6
|
|
; SI-NEXT: v_writelane_b32 v28, s39, 7
|
|
; SI-NEXT: v_writelane_b32 v28, s48, 8
|
|
; SI-NEXT: v_writelane_b32 v28, s49, 9
|
|
; SI-NEXT: v_writelane_b32 v28, s50, 10
|
|
; SI-NEXT: v_writelane_b32 v28, s51, 11
|
|
; SI-NEXT: v_writelane_b32 v28, s52, 12
|
|
; SI-NEXT: v_writelane_b32 v28, s53, 13
|
|
; SI-NEXT: v_writelane_b32 v28, s54, 14
|
|
; SI-NEXT: v_writelane_b32 v28, s55, 15
|
|
; SI-NEXT: v_writelane_b32 v28, s64, 16
|
|
; SI-NEXT: v_writelane_b32 v28, s65, 17
|
|
; SI-NEXT: v_writelane_b32 v28, s66, 18
|
|
; SI-NEXT: v_writelane_b32 v28, s67, 19
|
|
; SI-NEXT: v_writelane_b32 v28, s68, 20
|
|
; SI-NEXT: v_writelane_b32 v28, s69, 21
|
|
; SI-NEXT: v_writelane_b32 v28, s70, 22
|
|
; SI-NEXT: v_writelane_b32 v28, s71, 23
|
|
; SI-NEXT: v_writelane_b32 v28, s80, 24
|
|
; SI-NEXT: v_writelane_b32 v28, s81, 25
|
|
; SI-NEXT: v_writelane_b32 v28, s82, 26
|
|
; SI-NEXT: v_writelane_b32 v28, s83, 27
|
|
; SI-NEXT: v_writelane_b32 v28, s84, 28
|
|
; SI-NEXT: v_writelane_b32 v28, s85, 29
|
|
; SI-NEXT: v_writelane_b32 v28, s86, 30
|
|
; SI-NEXT: v_writelane_b32 v28, s87, 31
|
|
; SI-NEXT: v_writelane_b32 v28, s96, 32
|
|
; SI-NEXT: v_writelane_b32 v28, s97, 33
|
|
; SI-NEXT: v_writelane_b32 v28, s98, 34
|
|
; SI-NEXT: v_writelane_b32 v28, s99, 35
|
|
; SI-NEXT: v_readfirstlane_b32 s54, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s55, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s52, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s53, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s86, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s98, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s96, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s83, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s85, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s87, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s82, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s84, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s97, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s99, v0
|
|
; SI-NEXT: s_lshr_b32 s64, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s37, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s65, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s95, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s79, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s75, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s92, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s73, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s77, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s78, s54, 16
|
|
; SI-NEXT: s_lshr_b32 s80, s55, 16
|
|
; SI-NEXT: s_lshr_b32 s88, s52, 16
|
|
; SI-NEXT: s_lshr_b32 s91, s53, 16
|
|
; SI-NEXT: s_lshr_b32 s81, s86, 16
|
|
; SI-NEXT: s_lshr_b32 s69, s98, 16
|
|
; SI-NEXT: s_lshr_b32 s70, s96, 16
|
|
; SI-NEXT: s_lshr_b32 s67, s83, 16
|
|
; SI-NEXT: s_lshr_b32 s71, s85, 16
|
|
; SI-NEXT: s_lshr_b32 s51, s87, 16
|
|
; SI-NEXT: s_lshr_b32 s68, s82, 16
|
|
; SI-NEXT: s_lshr_b32 s49, s84, 16
|
|
; SI-NEXT: s_lshr_b32 s66, s97, 16
|
|
; SI-NEXT: s_lshr_b32 s39, s99, 16
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: ; implicit-def: $vgpr29 : SGPR spill to VGPR lane
|
|
; SI-NEXT: s_cbranch_scc0 .LBB57_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_and_b32 s5, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s73, 16
|
|
; SI-NEXT: s_mov_b32 s76, s73
|
|
; SI-NEXT: s_or_b32 s73, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s75, 16
|
|
; SI-NEXT: s_lshl_b32 s72, s77, 16
|
|
; SI-NEXT: s_mov_b32 s90, s77
|
|
; SI-NEXT: s_mov_b32 s77, s75
|
|
; SI-NEXT: s_or_b32 s75, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s79, 16
|
|
; SI-NEXT: s_or_b32 s63, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s89, 16
|
|
; SI-NEXT: s_or_b32 s61, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s50, 16
|
|
; SI-NEXT: s_or_b32 s59, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s65, 16
|
|
; SI-NEXT: s_or_b32 s57, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s64, 16
|
|
; SI-NEXT: s_or_b32 s47, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s97, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s66, 16
|
|
; SI-NEXT: s_or_b32 s45, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s82, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s68, 16
|
|
; SI-NEXT: s_or_b32 s43, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s85, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s71, 16
|
|
; SI-NEXT: s_or_b32 s41, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s96, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s70, 16
|
|
; SI-NEXT: s_or_b32 s15, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s86, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s81, 16
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_or_b32 s13, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s52, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s88, 16
|
|
; SI-NEXT: s_or_b32 s8, s4, s72
|
|
; SI-NEXT: s_and_b32 s4, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s74, s92, 16
|
|
; SI-NEXT: s_or_b32 s11, s5, s7
|
|
; SI-NEXT: s_and_b32 s5, s54, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s78, 16
|
|
; SI-NEXT: s_mov_b32 s9, s73
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[72:73], 16
|
|
; SI-NEXT: s_or_b32 s6, s4, s74
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s62, s93, 16
|
|
; SI-NEXT: s_or_b32 vcc_hi, s5, s7
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v29, s72, 0
|
|
; SI-NEXT: s_mov_b32 s7, s75
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[74:75], 16
|
|
; SI-NEXT: s_or_b32 s4, s4, s62
|
|
; SI-NEXT: s_lshl_b32 s60, s95, 16
|
|
; SI-NEXT: v_writelane_b32 v29, s73, 1
|
|
; SI-NEXT: s_mov_b32 s73, s76
|
|
; SI-NEXT: s_mov_b32 s75, s77
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[62:63], 16
|
|
; SI-NEXT: s_and_b32 s62, s22, 0xffff
|
|
; SI-NEXT: s_mov_b32 s5, s63
|
|
; SI-NEXT: s_or_b32 s62, s62, s60
|
|
; SI-NEXT: s_mov_b32 s63, s61
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[60:61], 16
|
|
; SI-NEXT: v_writelane_b32 v29, s60, 2
|
|
; SI-NEXT: s_lshl_b32 s58, s31, 16
|
|
; SI-NEXT: v_writelane_b32 v29, s61, 3
|
|
; SI-NEXT: s_and_b32 s60, s24, 0xffff
|
|
; SI-NEXT: s_or_b32 s60, s60, s58
|
|
; SI-NEXT: s_mov_b32 s61, s59
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[58:59], 16
|
|
; SI-NEXT: v_writelane_b32 v29, s58, 4
|
|
; SI-NEXT: s_lshl_b32 s56, s35, 16
|
|
; SI-NEXT: v_writelane_b32 v29, s59, 5
|
|
; SI-NEXT: s_and_b32 s58, s26, 0xffff
|
|
; SI-NEXT: s_or_b32 s58, s58, s56
|
|
; SI-NEXT: s_mov_b32 s59, s57
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[56:57], 16
|
|
; SI-NEXT: v_writelane_b32 v29, s56, 6
|
|
; SI-NEXT: s_lshl_b32 s46, s37, 16
|
|
; SI-NEXT: v_writelane_b32 v29, s57, 7
|
|
; SI-NEXT: s_and_b32 s56, s28, 0xffff
|
|
; SI-NEXT: s_or_b32 s56, s56, s46
|
|
; SI-NEXT: s_mov_b32 s57, s47
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[46:47], 16
|
|
; SI-NEXT: s_lshl_b32 s44, s39, 16
|
|
; SI-NEXT: v_writelane_b32 v29, s46, 8
|
|
; SI-NEXT: s_lshl_b32 s42, s49, 16
|
|
; SI-NEXT: v_writelane_b32 v29, s47, 9
|
|
; SI-NEXT: s_and_b32 s46, s99, 0xffff
|
|
; SI-NEXT: s_mov_b32 s30, s95
|
|
; SI-NEXT: s_lshr_b64 s[94:95], s[44:45], 16
|
|
; SI-NEXT: s_lshl_b32 s40, s51, 16
|
|
; SI-NEXT: s_or_b32 s46, s46, s44
|
|
; SI-NEXT: s_mov_b32 s95, s30
|
|
; SI-NEXT: s_and_b32 s44, s84, 0xffff
|
|
; SI-NEXT: s_mov_b32 s34, s31
|
|
; SI-NEXT: s_lshr_b64 s[30:31], s[42:43], 16
|
|
; SI-NEXT: s_lshl_b32 s14, s67, 16
|
|
; SI-NEXT: s_or_b32 s44, s44, s42
|
|
; SI-NEXT: s_mov_b32 s31, s34
|
|
; SI-NEXT: s_and_b32 s42, s87, 0xffff
|
|
; SI-NEXT: s_mov_b32 s36, s35
|
|
; SI-NEXT: s_lshr_b64 s[34:35], s[40:41], 16
|
|
; SI-NEXT: s_lshl_b32 s12, s69, 16
|
|
; SI-NEXT: s_or_b32 s42, s42, s40
|
|
; SI-NEXT: s_mov_b32 s35, s36
|
|
; SI-NEXT: s_and_b32 s40, s83, 0xffff
|
|
; SI-NEXT: s_mov_b32 s38, s37
|
|
; SI-NEXT: s_lshr_b64 s[36:37], s[14:15], 16
|
|
; SI-NEXT: s_lshl_b32 s10, s91, 16
|
|
; SI-NEXT: s_or_b32 s40, s40, s14
|
|
; SI-NEXT: s_mov_b32 s37, s38
|
|
; SI-NEXT: s_and_b32 s14, s98, 0xffff
|
|
; SI-NEXT: s_mov_b32 s48, s39
|
|
; SI-NEXT: s_lshr_b64 s[38:39], s[12:13], 16
|
|
; SI-NEXT: s_lshl_b32 vcc_lo, s80, 16
|
|
; SI-NEXT: s_mov_b32 s72, s80
|
|
; SI-NEXT: s_mov_b32 s80, s70
|
|
; SI-NEXT: s_mov_b32 s70, s68
|
|
; SI-NEXT: s_mov_b32 s68, s66
|
|
; SI-NEXT: s_mov_b32 s66, s19
|
|
; SI-NEXT: s_or_b32 s14, s14, s12
|
|
; SI-NEXT: s_mov_b32 s39, s48
|
|
; SI-NEXT: s_and_b32 s12, s53, 0xffff
|
|
; SI-NEXT: s_mov_b32 s19, s64
|
|
; SI-NEXT: s_mov_b32 s64, s50
|
|
; SI-NEXT: s_mov_b32 s50, s49
|
|
; SI-NEXT: s_lshr_b64 s[48:49], s[10:11], 16
|
|
; SI-NEXT: s_mov_b32 s77, s90
|
|
; SI-NEXT: s_or_b32 s12, s12, s10
|
|
; SI-NEXT: s_mov_b32 s49, s50
|
|
; SI-NEXT: s_mov_b32 s50, s64
|
|
; SI-NEXT: s_mov_b32 s64, s19
|
|
; SI-NEXT: s_mov_b32 s19, s66
|
|
; SI-NEXT: s_mov_b32 s66, s68
|
|
; SI-NEXT: s_mov_b32 s68, s70
|
|
; SI-NEXT: s_mov_b32 s70, s80
|
|
; SI-NEXT: s_and_b32 s10, s55, 0xffff
|
|
; SI-NEXT: s_mov_b32 s80, s91
|
|
; SI-NEXT: s_lshr_b64 s[90:91], vcc, 16
|
|
; SI-NEXT: s_mov_b32 s47, s45
|
|
; SI-NEXT: s_mov_b32 s45, s43
|
|
; SI-NEXT: s_mov_b32 s43, s41
|
|
; SI-NEXT: s_mov_b32 s41, s15
|
|
; SI-NEXT: s_mov_b32 s15, s13
|
|
; SI-NEXT: s_mov_b32 s13, s11
|
|
; SI-NEXT: s_or_b32 s10, s10, vcc_lo
|
|
; SI-NEXT: s_mov_b32 s11, vcc_hi
|
|
; SI-NEXT: s_mov_b32 s91, s80
|
|
; SI-NEXT: s_mov_b32 s80, s72
|
|
; SI-NEXT: s_cbranch_execnz .LBB57_3
|
|
; SI-NEXT: .LBB57_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s55, s55, 3
|
|
; SI-NEXT: s_and_b32 s4, s55, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s80, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s54, s54, 3
|
|
; SI-NEXT: s_add_i32 s10, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s54, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s78, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s53, s53, 3
|
|
; SI-NEXT: s_add_i32 s11, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s53, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s91, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s52, s52, 3
|
|
; SI-NEXT: s_add_i32 s12, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s52, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s88, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s98, s98, 3
|
|
; SI-NEXT: s_add_i32 s13, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s98, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s69, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s86, s86, 3
|
|
; SI-NEXT: s_add_i32 s14, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s86, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s81, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s83, s83, 3
|
|
; SI-NEXT: s_add_i32 s15, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s83, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s67, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s96, s96, 3
|
|
; SI-NEXT: s_add_i32 s40, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s96, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s70, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s87, s87, 3
|
|
; SI-NEXT: s_add_i32 s41, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s87, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s51, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s85, s85, 3
|
|
; SI-NEXT: s_add_i32 s42, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s85, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s71, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s84, s84, 3
|
|
; SI-NEXT: s_add_i32 s43, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s84, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s49, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s82, s82, 3
|
|
; SI-NEXT: s_add_i32 s44, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s82, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s68, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s99, s99, 3
|
|
; SI-NEXT: s_add_i32 s45, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s99, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s39, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s97, s97, 3
|
|
; SI-NEXT: s_add_i32 s46, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s97, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s66, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s47, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s37, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: s_add_i32 s56, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s29, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s64, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s57, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s58, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s27, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s65, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s59, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s31, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s60, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s50, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: s_add_i32 s61, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s95, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s62, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s89, 16
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: s_add_i32 s63, s4, 0x30000
|
|
; SI-NEXT: s_and_b32 s4, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s93, 16
|
|
; SI-NEXT: s_add_i32 s21, s21, 3
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_and_b32 s5, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s79, 16
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: s_or_b32 s5, s6, s5
|
|
; SI-NEXT: s_and_b32 s6, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s92, 16
|
|
; SI-NEXT: s_add_i32 s19, s19, 3
|
|
; SI-NEXT: s_or_b32 s6, s7, s6
|
|
; SI-NEXT: s_and_b32 s7, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s75, 16
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: s_or_b32 s7, s8, s7
|
|
; SI-NEXT: s_and_b32 s8, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s77, 16
|
|
; SI-NEXT: s_add_i32 s17, s17, 3
|
|
; SI-NEXT: s_or_b32 s8, s9, s8
|
|
; SI-NEXT: s_and_b32 s9, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s16, s73, 16
|
|
; SI-NEXT: s_or_b32 s9, s16, s9
|
|
; SI-NEXT: s_add_i32 s8, s8, 0x30000
|
|
; SI-NEXT: s_add_i32 s9, s9, 0x30000
|
|
; SI-NEXT: s_lshr_b64 s[16:17], s[8:9], 16
|
|
; SI-NEXT: v_writelane_b32 v29, s16, 0
|
|
; SI-NEXT: v_writelane_b32 v29, s17, 1
|
|
; SI-NEXT: s_lshr_b64 s[16:17], s[62:63], 16
|
|
; SI-NEXT: v_writelane_b32 v29, s16, 2
|
|
; SI-NEXT: v_writelane_b32 v29, s17, 3
|
|
; SI-NEXT: s_lshr_b64 s[16:17], s[60:61], 16
|
|
; SI-NEXT: v_writelane_b32 v29, s16, 4
|
|
; SI-NEXT: v_writelane_b32 v29, s17, 5
|
|
; SI-NEXT: s_lshr_b64 s[16:17], s[58:59], 16
|
|
; SI-NEXT: s_add_i32 s6, s6, 0x30000
|
|
; SI-NEXT: s_add_i32 s7, s7, 0x30000
|
|
; SI-NEXT: v_writelane_b32 v29, s16, 6
|
|
; SI-NEXT: s_add_i32 s4, s4, 0x30000
|
|
; SI-NEXT: s_add_i32 s5, s5, 0x30000
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[6:7], 16
|
|
; SI-NEXT: v_writelane_b32 v29, s17, 7
|
|
; SI-NEXT: s_lshr_b64 s[16:17], s[56:57], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[4:5], 16
|
|
; SI-NEXT: v_writelane_b32 v29, s16, 8
|
|
; SI-NEXT: s_lshr_b64 s[94:95], s[46:47], 16
|
|
; SI-NEXT: s_lshr_b64 s[30:31], s[44:45], 16
|
|
; SI-NEXT: s_lshr_b64 s[34:35], s[42:43], 16
|
|
; SI-NEXT: s_lshr_b64 s[36:37], s[40:41], 16
|
|
; SI-NEXT: s_lshr_b64 s[38:39], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[48:49], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[90:91], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b32 s73, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s75, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s79, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s63, 16
|
|
; SI-NEXT: s_lshr_b32 s50, s61, 16
|
|
; SI-NEXT: s_lshr_b32 s65, s59, 16
|
|
; SI-NEXT: s_lshr_b32 s64, s57, 16
|
|
; SI-NEXT: s_lshr_b32 s66, s47, 16
|
|
; SI-NEXT: s_lshr_b32 s68, s45, 16
|
|
; SI-NEXT: s_lshr_b32 s71, s43, 16
|
|
; SI-NEXT: s_lshr_b32 s70, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s81, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s88, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s78, s11, 16
|
|
; SI-NEXT: v_writelane_b32 v29, s17, 9
|
|
; SI-NEXT: .LBB57_3: ; %end
|
|
; SI-NEXT: v_readlane_b32 s16, v29, 0
|
|
; SI-NEXT: s_and_b32 s8, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s16, s16, 16
|
|
; SI-NEXT: s_or_b32 s8, s8, s16
|
|
; SI-NEXT: s_and_b32 s9, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s16, s73, 16
|
|
; SI-NEXT: s_or_b32 s9, s9, s16
|
|
; SI-NEXT: s_and_b32 s6, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s16, s74, 16
|
|
; SI-NEXT: s_or_b32 s6, s6, s16
|
|
; SI-NEXT: s_and_b32 s7, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s16, s75, 16
|
|
; SI-NEXT: s_or_b32 s7, s7, s16
|
|
; SI-NEXT: s_and_b32 s4, s4, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s16, s76, 16
|
|
; SI-NEXT: v_readlane_b32 s17, v29, 1
|
|
; SI-NEXT: s_or_b32 s4, s4, s16
|
|
; SI-NEXT: s_and_b32 s5, s5, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s16, s79, 16
|
|
; SI-NEXT: v_readlane_b32 s18, v29, 2
|
|
; SI-NEXT: s_or_b32 s5, s5, s16
|
|
; SI-NEXT: s_and_b32 s16, s62, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s17, s18, 16
|
|
; SI-NEXT: v_readlane_b32 s19, v29, 3
|
|
; SI-NEXT: s_or_b32 s16, s16, s17
|
|
; SI-NEXT: s_and_b32 s17, s63, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s18, s89, 16
|
|
; SI-NEXT: v_readlane_b32 s20, v29, 4
|
|
; SI-NEXT: s_or_b32 s17, s17, s18
|
|
; SI-NEXT: s_and_b32 s18, s60, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s19, s20, 16
|
|
; SI-NEXT: v_readlane_b32 s21, v29, 5
|
|
; SI-NEXT: s_or_b32 s18, s18, s19
|
|
; SI-NEXT: s_and_b32 s19, s61, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s20, s50, 16
|
|
; SI-NEXT: v_readlane_b32 s22, v29, 6
|
|
; SI-NEXT: s_or_b32 s19, s19, s20
|
|
; SI-NEXT: s_and_b32 s20, s58, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s21, s22, 16
|
|
; SI-NEXT: v_readlane_b32 s23, v29, 7
|
|
; SI-NEXT: s_or_b32 s20, s20, s21
|
|
; SI-NEXT: s_and_b32 s21, s59, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s22, s65, 16
|
|
; SI-NEXT: v_readlane_b32 s24, v29, 8
|
|
; SI-NEXT: s_or_b32 s21, s21, s22
|
|
; SI-NEXT: s_and_b32 s22, s56, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s23, s24, 16
|
|
; SI-NEXT: v_readlane_b32 s25, v29, 9
|
|
; SI-NEXT: s_or_b32 s22, s22, s23
|
|
; SI-NEXT: s_and_b32 s23, s57, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s24, s64, 16
|
|
; SI-NEXT: s_or_b32 s23, s23, s24
|
|
; SI-NEXT: s_and_b32 s24, s46, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s25, s94, 16
|
|
; SI-NEXT: s_or_b32 s24, s24, s25
|
|
; SI-NEXT: s_and_b32 s25, s47, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s26, s66, 16
|
|
; SI-NEXT: s_or_b32 s25, s25, s26
|
|
; SI-NEXT: s_and_b32 s26, s44, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s27, s30, 16
|
|
; SI-NEXT: s_or_b32 s26, s26, s27
|
|
; SI-NEXT: s_and_b32 s27, s45, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s28, s68, 16
|
|
; SI-NEXT: s_or_b32 s27, s27, s28
|
|
; SI-NEXT: s_and_b32 s28, s42, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s29, s34, 16
|
|
; SI-NEXT: s_or_b32 s28, s28, s29
|
|
; SI-NEXT: s_and_b32 s29, s43, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s71, 16
|
|
; SI-NEXT: s_or_b32 s29, s29, s42
|
|
; SI-NEXT: s_and_b32 s40, s40, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s36, 16
|
|
; SI-NEXT: s_or_b32 s40, s40, s42
|
|
; SI-NEXT: s_and_b32 s41, s41, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s70, 16
|
|
; SI-NEXT: s_or_b32 s41, s41, s42
|
|
; SI-NEXT: s_and_b32 s14, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s38, 16
|
|
; SI-NEXT: s_or_b32 s14, s14, s42
|
|
; SI-NEXT: s_and_b32 s15, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s81, 16
|
|
; SI-NEXT: s_or_b32 s15, s15, s42
|
|
; SI-NEXT: s_and_b32 s12, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s48, 16
|
|
; SI-NEXT: s_or_b32 s12, s12, s42
|
|
; SI-NEXT: s_and_b32 s13, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s88, 16
|
|
; SI-NEXT: s_or_b32 s13, s13, s42
|
|
; SI-NEXT: s_and_b32 s10, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s90, 16
|
|
; SI-NEXT: s_or_b32 s10, s10, s42
|
|
; SI-NEXT: s_and_b32 s11, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s42, s78, 16
|
|
; SI-NEXT: s_or_b32 s11, s11, s42
|
|
; SI-NEXT: v_mov_b32_e32 v0, s8
|
|
; SI-NEXT: v_mov_b32_e32 v1, s9
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: v_mov_b32_e32 v3, s7
|
|
; SI-NEXT: v_mov_b32_e32 v4, s4
|
|
; SI-NEXT: v_mov_b32_e32 v5, s5
|
|
; SI-NEXT: v_mov_b32_e32 v6, s16
|
|
; SI-NEXT: v_mov_b32_e32 v7, s17
|
|
; SI-NEXT: v_mov_b32_e32 v8, s18
|
|
; SI-NEXT: v_mov_b32_e32 v9, s19
|
|
; SI-NEXT: v_mov_b32_e32 v10, s20
|
|
; SI-NEXT: v_mov_b32_e32 v11, s21
|
|
; SI-NEXT: v_mov_b32_e32 v12, s22
|
|
; SI-NEXT: v_mov_b32_e32 v13, s23
|
|
; SI-NEXT: v_mov_b32_e32 v14, s24
|
|
; SI-NEXT: v_mov_b32_e32 v15, s25
|
|
; SI-NEXT: v_mov_b32_e32 v16, s26
|
|
; SI-NEXT: v_mov_b32_e32 v17, s27
|
|
; SI-NEXT: v_mov_b32_e32 v18, s28
|
|
; SI-NEXT: v_mov_b32_e32 v19, s29
|
|
; SI-NEXT: v_mov_b32_e32 v20, s40
|
|
; SI-NEXT: v_mov_b32_e32 v21, s41
|
|
; SI-NEXT: v_mov_b32_e32 v22, s14
|
|
; SI-NEXT: v_mov_b32_e32 v23, s15
|
|
; SI-NEXT: v_mov_b32_e32 v24, s12
|
|
; SI-NEXT: v_mov_b32_e32 v25, s13
|
|
; SI-NEXT: v_mov_b32_e32 v26, s10
|
|
; SI-NEXT: v_mov_b32_e32 v27, s11
|
|
; SI-NEXT: v_readlane_b32 s99, v28, 35
|
|
; SI-NEXT: v_readlane_b32 s98, v28, 34
|
|
; SI-NEXT: v_readlane_b32 s97, v28, 33
|
|
; SI-NEXT: v_readlane_b32 s96, v28, 32
|
|
; SI-NEXT: v_readlane_b32 s87, v28, 31
|
|
; SI-NEXT: v_readlane_b32 s86, v28, 30
|
|
; SI-NEXT: v_readlane_b32 s85, v28, 29
|
|
; SI-NEXT: v_readlane_b32 s84, v28, 28
|
|
; SI-NEXT: v_readlane_b32 s83, v28, 27
|
|
; SI-NEXT: v_readlane_b32 s82, v28, 26
|
|
; SI-NEXT: v_readlane_b32 s81, v28, 25
|
|
; SI-NEXT: v_readlane_b32 s80, v28, 24
|
|
; SI-NEXT: v_readlane_b32 s71, v28, 23
|
|
; SI-NEXT: v_readlane_b32 s70, v28, 22
|
|
; SI-NEXT: v_readlane_b32 s69, v28, 21
|
|
; SI-NEXT: v_readlane_b32 s68, v28, 20
|
|
; SI-NEXT: v_readlane_b32 s67, v28, 19
|
|
; SI-NEXT: v_readlane_b32 s66, v28, 18
|
|
; SI-NEXT: v_readlane_b32 s65, v28, 17
|
|
; SI-NEXT: v_readlane_b32 s64, v28, 16
|
|
; SI-NEXT: v_readlane_b32 s55, v28, 15
|
|
; SI-NEXT: v_readlane_b32 s54, v28, 14
|
|
; SI-NEXT: v_readlane_b32 s53, v28, 13
|
|
; SI-NEXT: v_readlane_b32 s52, v28, 12
|
|
; SI-NEXT: v_readlane_b32 s51, v28, 11
|
|
; SI-NEXT: v_readlane_b32 s50, v28, 10
|
|
; SI-NEXT: v_readlane_b32 s49, v28, 9
|
|
; SI-NEXT: v_readlane_b32 s48, v28, 8
|
|
; SI-NEXT: v_readlane_b32 s39, v28, 7
|
|
; SI-NEXT: v_readlane_b32 s38, v28, 6
|
|
; SI-NEXT: v_readlane_b32 s37, v28, 5
|
|
; SI-NEXT: v_readlane_b32 s36, v28, 4
|
|
; SI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB57_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr4
|
|
; SI-NEXT: ; implicit-def: $sgpr10
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_writelane_b32 v29, s4, 0
|
|
; SI-NEXT: v_writelane_b32 v29, s5, 1
|
|
; SI-NEXT: v_writelane_b32 v29, s10, 2
|
|
; SI-NEXT: v_writelane_b32 v29, s11, 3
|
|
; SI-NEXT: ; implicit-def: $sgpr10
|
|
; SI-NEXT: ; implicit-def: $sgpr8
|
|
; SI-NEXT: ; implicit-def: $sgpr6
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr4
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr94
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr40
|
|
; SI-NEXT: ; implicit-def: $sgpr36
|
|
; SI-NEXT: ; implicit-def: $sgpr14
|
|
; SI-NEXT: ; implicit-def: $sgpr38
|
|
; SI-NEXT: ; implicit-def: $sgpr12
|
|
; SI-NEXT: ; implicit-def: $sgpr48
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: v_writelane_b32 v29, s10, 4
|
|
; SI-NEXT: v_writelane_b32 v29, s11, 5
|
|
; SI-NEXT: ; implicit-def: $sgpr10
|
|
; SI-NEXT: v_writelane_b32 v29, s10, 6
|
|
; SI-NEXT: v_writelane_b32 v29, s11, 7
|
|
; SI-NEXT: ; implicit-def: $sgpr10
|
|
; SI-NEXT: v_writelane_b32 v29, s10, 8
|
|
; SI-NEXT: v_writelane_b32 v29, s11, 9
|
|
; SI-NEXT: ; implicit-def: $sgpr10
|
|
; SI-NEXT: s_branch .LBB57_2
|
|
;
|
|
; VI-LABEL: bitcast_v56i16_to_v56f16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v28, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v28, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v28, s31, 1
|
|
; VI-NEXT: v_writelane_b32 v28, s34, 2
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s43, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s47, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s58, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s61, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s72, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s76, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s79, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s90, v0
|
|
; VI-NEXT: v_writelane_b32 v28, s35, 3
|
|
; VI-NEXT: s_lshr_b32 s42, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s30, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s34, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s6, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s9, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s11, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s13, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s15, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s41, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s44, s43, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s47, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s58, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s61, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s72, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s76, 16
|
|
; VI-NEXT: s_lshr_b32 s88, s79, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s90, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: s_cbranch_scc0 .LBB57_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB57_3
|
|
; VI-NEXT: .LBB57_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: s_add_i32 s35, s35, 3
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_add_i32 s34, s34, 3
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s31, s31, 3
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s30, s30, 3
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s89, s89, 3
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s78, s78, 3
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s75, s75, 3
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s74, s74, 3
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s63, s63, 3
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s60, s60, 3
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s57, s57, 3
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s46, s46, 3
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s45, s45, 3
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s42, s42, 3
|
|
; VI-NEXT: s_add_i32 s90, s90, 3
|
|
; VI-NEXT: s_add_i32 s91, s91, 3
|
|
; VI-NEXT: s_add_i32 s79, s79, 3
|
|
; VI-NEXT: s_add_i32 s88, s88, 3
|
|
; VI-NEXT: s_add_i32 s76, s76, 3
|
|
; VI-NEXT: s_add_i32 s77, s77, 3
|
|
; VI-NEXT: s_add_i32 s72, s72, 3
|
|
; VI-NEXT: s_add_i32 s73, s73, 3
|
|
; VI-NEXT: s_add_i32 s61, s61, 3
|
|
; VI-NEXT: s_add_i32 s62, s62, 3
|
|
; VI-NEXT: s_add_i32 s58, s58, 3
|
|
; VI-NEXT: s_add_i32 s59, s59, 3
|
|
; VI-NEXT: s_add_i32 s47, s47, 3
|
|
; VI-NEXT: s_add_i32 s56, s56, 3
|
|
; VI-NEXT: s_add_i32 s43, s43, 3
|
|
; VI-NEXT: s_add_i32 s44, s44, 3
|
|
; VI-NEXT: s_add_i32 s40, s40, 3
|
|
; VI-NEXT: s_add_i32 s41, s41, 3
|
|
; VI-NEXT: s_add_i32 s14, s14, 3
|
|
; VI-NEXT: s_add_i32 s15, s15, 3
|
|
; VI-NEXT: s_add_i32 s12, s12, 3
|
|
; VI-NEXT: s_add_i32 s13, s13, 3
|
|
; VI-NEXT: s_add_i32 s10, s10, 3
|
|
; VI-NEXT: s_add_i32 s11, s11, 3
|
|
; VI-NEXT: s_add_i32 s8, s8, 3
|
|
; VI-NEXT: s_add_i32 s9, s9, 3
|
|
; VI-NEXT: s_add_i32 s7, s7, 3
|
|
; VI-NEXT: s_add_i32 s6, s6, 3
|
|
; VI-NEXT: .LBB57_3: ; %end
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s35, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s16, s34, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s16
|
|
; VI-NEXT: s_and_b32 s16, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s17, s31, 16
|
|
; VI-NEXT: s_or_b32 s16, s16, s17
|
|
; VI-NEXT: s_and_b32 s17, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s18, s30, 16
|
|
; VI-NEXT: s_or_b32 s17, s17, s18
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s19, s89, 16
|
|
; VI-NEXT: s_or_b32 s18, s18, s19
|
|
; VI-NEXT: s_and_b32 s19, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s20, s78, 16
|
|
; VI-NEXT: s_or_b32 s19, s19, s20
|
|
; VI-NEXT: s_and_b32 s20, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s21, s75, 16
|
|
; VI-NEXT: s_or_b32 s20, s20, s21
|
|
; VI-NEXT: s_and_b32 s21, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s22, s74, 16
|
|
; VI-NEXT: s_or_b32 s21, s21, s22
|
|
; VI-NEXT: s_and_b32 s22, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s23, s63, 16
|
|
; VI-NEXT: s_or_b32 s22, s22, s23
|
|
; VI-NEXT: s_and_b32 s23, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s24, s60, 16
|
|
; VI-NEXT: s_or_b32 s23, s23, s24
|
|
; VI-NEXT: s_and_b32 s24, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s25, s57, 16
|
|
; VI-NEXT: s_or_b32 s24, s24, s25
|
|
; VI-NEXT: s_and_b32 s25, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s26, s46, 16
|
|
; VI-NEXT: s_or_b32 s25, s25, s26
|
|
; VI-NEXT: s_and_b32 s26, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s27, s45, 16
|
|
; VI-NEXT: s_or_b32 s26, s26, s27
|
|
; VI-NEXT: s_and_b32 s27, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s28, s42, 16
|
|
; VI-NEXT: s_or_b32 s27, s27, s28
|
|
; VI-NEXT: s_and_b32 s28, 0xffff, s90
|
|
; VI-NEXT: s_lshl_b32 s29, s91, 16
|
|
; VI-NEXT: s_or_b32 s28, s28, s29
|
|
; VI-NEXT: s_and_b32 s29, 0xffff, s79
|
|
; VI-NEXT: s_lshl_b32 s42, s88, 16
|
|
; VI-NEXT: s_or_b32 s29, s29, s42
|
|
; VI-NEXT: s_and_b32 s42, 0xffff, s76
|
|
; VI-NEXT: s_lshl_b32 s45, s77, 16
|
|
; VI-NEXT: s_or_b32 s42, s42, s45
|
|
; VI-NEXT: s_and_b32 s45, 0xffff, s72
|
|
; VI-NEXT: s_lshl_b32 s46, s73, 16
|
|
; VI-NEXT: s_or_b32 s45, s45, s46
|
|
; VI-NEXT: s_and_b32 s46, 0xffff, s61
|
|
; VI-NEXT: s_lshl_b32 s57, s62, 16
|
|
; VI-NEXT: s_or_b32 s46, s46, s57
|
|
; VI-NEXT: s_and_b32 s57, 0xffff, s58
|
|
; VI-NEXT: s_lshl_b32 s58, s59, 16
|
|
; VI-NEXT: s_and_b32 s47, 0xffff, s47
|
|
; VI-NEXT: s_lshl_b32 s56, s56, 16
|
|
; VI-NEXT: s_and_b32 s43, 0xffff, s43
|
|
; VI-NEXT: s_lshl_b32 s44, s44, 16
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s40
|
|
; VI-NEXT: s_lshl_b32 s41, s41, 16
|
|
; VI-NEXT: s_and_b32 s14, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s15, s15, 16
|
|
; VI-NEXT: s_and_b32 s12, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s13, s13, 16
|
|
; VI-NEXT: s_and_b32 s10, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s11, s11, 16
|
|
; VI-NEXT: s_and_b32 s8, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s9, s9, 16
|
|
; VI-NEXT: s_and_b32 s7, 0xffff, s7
|
|
; VI-NEXT: s_lshl_b32 s6, s6, 16
|
|
; VI-NEXT: s_or_b32 s57, s57, s58
|
|
; VI-NEXT: s_or_b32 s47, s47, s56
|
|
; VI-NEXT: s_or_b32 s43, s43, s44
|
|
; VI-NEXT: s_or_b32 s40, s40, s41
|
|
; VI-NEXT: s_or_b32 s14, s14, s15
|
|
; VI-NEXT: s_or_b32 s12, s12, s13
|
|
; VI-NEXT: s_or_b32 s10, s10, s11
|
|
; VI-NEXT: s_or_b32 s8, s8, s9
|
|
; VI-NEXT: s_or_b32 s6, s7, s6
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s16
|
|
; VI-NEXT: v_mov_b32_e32 v3, s17
|
|
; VI-NEXT: v_mov_b32_e32 v4, s18
|
|
; VI-NEXT: v_mov_b32_e32 v5, s19
|
|
; VI-NEXT: v_mov_b32_e32 v6, s20
|
|
; VI-NEXT: v_mov_b32_e32 v7, s21
|
|
; VI-NEXT: v_mov_b32_e32 v8, s22
|
|
; VI-NEXT: v_mov_b32_e32 v9, s23
|
|
; VI-NEXT: v_mov_b32_e32 v10, s24
|
|
; VI-NEXT: v_mov_b32_e32 v11, s25
|
|
; VI-NEXT: v_mov_b32_e32 v12, s26
|
|
; VI-NEXT: v_mov_b32_e32 v13, s27
|
|
; VI-NEXT: v_mov_b32_e32 v14, s28
|
|
; VI-NEXT: v_mov_b32_e32 v15, s29
|
|
; VI-NEXT: v_mov_b32_e32 v16, s42
|
|
; VI-NEXT: v_mov_b32_e32 v17, s45
|
|
; VI-NEXT: v_mov_b32_e32 v18, s46
|
|
; VI-NEXT: v_mov_b32_e32 v19, s57
|
|
; VI-NEXT: v_mov_b32_e32 v20, s47
|
|
; VI-NEXT: v_mov_b32_e32 v21, s43
|
|
; VI-NEXT: v_mov_b32_e32 v22, s40
|
|
; VI-NEXT: v_mov_b32_e32 v23, s14
|
|
; VI-NEXT: v_mov_b32_e32 v24, s12
|
|
; VI-NEXT: v_mov_b32_e32 v25, s10
|
|
; VI-NEXT: v_mov_b32_e32 v26, s8
|
|
; VI-NEXT: v_mov_b32_e32 v27, s6
|
|
; VI-NEXT: v_readlane_b32 s35, v28, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v28, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v28, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v28, 0
|
|
; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v28, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB57_4:
|
|
; VI-NEXT: s_branch .LBB57_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v56i16_to_v56f16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s95, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s94, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s93, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s92, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s91, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s90, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s89, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s88, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s79, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s78, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s77, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s76, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s75, v1
|
|
; GFX9-NEXT: v_readfirstlane_b32 s74, v0
|
|
; GFX9-NEXT: s_lshr_b32 s43, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s40, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s15, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s14, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s13, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s12, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s11, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s10, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s9, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s8, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s7, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s6, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s95, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s94, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s93, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s92, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s91, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s90, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s89, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s88, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s79, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s78, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s77, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s76, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s75, 16
|
|
; GFX9-NEXT: s_lshr_b32 s44, s74, 16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB57_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB57_4
|
|
; GFX9-NEXT: .LBB57_2: ; %cmp.true
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s95, s73
|
|
; GFX9-NEXT: v_pk_add_u16 v27, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s94, s72
|
|
; GFX9-NEXT: v_pk_add_u16 v26, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s93, s63
|
|
; GFX9-NEXT: v_pk_add_u16 v25, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s92, s62
|
|
; GFX9-NEXT: v_pk_add_u16 v24, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s91, s61
|
|
; GFX9-NEXT: v_pk_add_u16 v23, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s90, s60
|
|
; GFX9-NEXT: v_pk_add_u16 v22, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s89, s59
|
|
; GFX9-NEXT: v_pk_add_u16 v21, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s88, s58
|
|
; GFX9-NEXT: v_pk_add_u16 v20, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s79, s57
|
|
; GFX9-NEXT: v_pk_add_u16 v19, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s78, s56
|
|
; GFX9-NEXT: v_pk_add_u16 v18, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s77, s47
|
|
; GFX9-NEXT: v_pk_add_u16 v17, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s76, s46
|
|
; GFX9-NEXT: v_pk_add_u16 v16, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s75, s45
|
|
; GFX9-NEXT: v_pk_add_u16 v15, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s74, s44
|
|
; GFX9-NEXT: v_pk_add_u16 v14, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s29, s43
|
|
; GFX9-NEXT: v_pk_add_u16 v13, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s28, s42
|
|
; GFX9-NEXT: v_pk_add_u16 v12, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s27, s41
|
|
; GFX9-NEXT: v_pk_add_u16 v11, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s26, s40
|
|
; GFX9-NEXT: v_pk_add_u16 v10, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s25, s15
|
|
; GFX9-NEXT: v_pk_add_u16 v9, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s24, s14
|
|
; GFX9-NEXT: v_pk_add_u16 v8, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s23, s13
|
|
; GFX9-NEXT: v_pk_add_u16 v7, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s22, s12
|
|
; GFX9-NEXT: v_pk_add_u16 v6, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s21, s11
|
|
; GFX9-NEXT: v_pk_add_u16 v5, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s20, s10
|
|
; GFX9-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s19, s9
|
|
; GFX9-NEXT: v_pk_add_u16 v3, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s18, s8
|
|
; GFX9-NEXT: v_pk_add_u16 v2, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s17, s7
|
|
; GFX9-NEXT: v_pk_add_u16 v1, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s6
|
|
; GFX9-NEXT: v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: s_branch .LBB57_5
|
|
; GFX9-NEXT: .LBB57_3:
|
|
; GFX9-NEXT: s_branch .LBB57_2
|
|
; GFX9-NEXT: .LBB57_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s95
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s94
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s93
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s92
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s91
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s90
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s89
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s88
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s79
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s78
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s77
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s76
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s75
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s74
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s73
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s72
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, s6
|
|
; GFX9-NEXT: .LBB57_5: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX9-NEXT: v_lshl_or_b32 v0, v47, 16, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v1, v46, 16, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v2, v45, 16, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v3, v44, 16, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v4, v43, 16, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v5, v42, 16, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v6, v41, 16, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v7, v40, 16, v7
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX9-NEXT: v_lshl_or_b32 v8, v55, 16, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v9, v54, 16, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v53, 16, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v52, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v51, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v50, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v49, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v48, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v39, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v38, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v37, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v36, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v35, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v34, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v33, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v32, 16, v23
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v31, 16, v24
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v30, 16, v25
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v29, 16, v26
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v28, 16, v27
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v56f16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s79, v9
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s89, v8
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s88, v7
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s78, v6
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s76, v5
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s74, v4
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s77, v3
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s75, v2
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s73, v1
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s72, v0
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s90, v10
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s59, s79, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s63, s89, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s62, s88, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s60, s78, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s57, s76, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s74, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s61, s77, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s58, s75, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s56, s73, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s72, 16
|
|
; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s90, 0
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s90, 0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB57_3
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s90
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB57_4
|
|
; GFX11-TRUE16-NEXT: .LBB57_2: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s59, s79, s59
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s57, s76, s57
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, s59, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s59, s88, s62
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s47, s74, s47
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s63, s89, s63
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, s59, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s59, s78, s60
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, s57, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s57, s77, s61
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, s47, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s47, s75, s58
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s56, s73, s56
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s46, s72, s46
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s29, s29, s45
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s28, s28, s44
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s27, s27, s43
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s26, s26, s42
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s25, s25, s41
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s24, s24, s40
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s23, s15
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s22, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s21, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s20, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s19, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s18, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s17, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s4
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, s63, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, s59, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, s47, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, s56, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, s46, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s29, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s28, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s26, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s25, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s24, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB57_5
|
|
; GFX11-TRUE16-NEXT: .LBB57_3:
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB57_2
|
|
; GFX11-TRUE16-NEXT: .LBB57_4:
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, s79 :: v_dual_mov_b32 v26, s89
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, s88 :: v_dual_mov_b32 v24, s78
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, s76 :: v_dual_mov_b32 v22, s74
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, s77 :: v_dual_mov_b32 v20, s75
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, s73 :: v_dual_mov_b32 v18, s72
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, s29 :: v_dual_mov_b32 v16, s28
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v14, s26
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v13, s25 :: v_dual_mov_b32 v12, s24
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v10, s22
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, s21 :: v_dual_mov_b32 v8, s20
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v6, s18
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v4, s16
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v0, s0
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, s59 :: v_dual_mov_b32 v29, s63
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, s62 :: v_dual_mov_b32 v31, s60
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v32, s57 :: v_dual_mov_b32 v33, s47
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v34, s61 :: v_dual_mov_b32 v35, s58
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v36, s56 :: v_dual_mov_b32 v37, s46
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v38, s45 :: v_dual_mov_b32 v39, s44
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v48, s43 :: v_dual_mov_b32 v49, s42
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v50, s41 :: v_dual_mov_b32 v51, s40
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v52, s15 :: v_dual_mov_b32 v53, s14
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s13 :: v_dual_mov_b32 v55, s12
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, s11 :: v_dual_mov_b32 v65, s10
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v66, s9 :: v_dual_mov_b32 v67, s6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v68, s8 :: v_dual_mov_b32 v69, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v70, s4 :: v_dual_mov_b32 v71, s5
|
|
; GFX11-TRUE16-NEXT: .LBB57_5: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v71.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v70.l
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v69.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v68.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v67.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v66.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v65.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v64.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v31.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v28.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56i16_to_v56f16_scalar:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s79, v9
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s89, v8
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s88, v7
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s78, v6
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s76, v5
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s74, v4
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s77, v3
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s75, v2
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s73, v1
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s72, v0
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s90, v10
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s59, s79, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s63, s89, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s62, s88, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s60, s78, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s57, s76, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s74, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s61, s77, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s58, s75, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s56, s73, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s72, 16
|
|
; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s90, 0
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s90, 0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB57_3
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s90
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB57_4
|
|
; GFX11-FAKE16-NEXT: .LBB57_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s59, s79, s59
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s57, s76, s57
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s59, s88, s62
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s47, s74, s47
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s63, s89, s63
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, s59, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s59, s78, s60
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v27, s57, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s57, s77, s61
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v18, s47, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s47, s75, s58
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s56, s73, s56
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s46, s72, s46
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s29, s29, s45
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s28, s28, s44
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s27, s27, s43
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s26, s26, s42
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s25, s25, s41
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s24, s24, s40
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s23, s15
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s22, s14
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s21, s13
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s20, s12
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s19, s11
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s18, s10
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s17, s9
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s8
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s0, s5
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s4
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, s63, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v26, s59, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v19, s57, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v20, s47, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v21, s56, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v22, s46, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, s29, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, s28, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, s26, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, s25, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, s24, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v23
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB57_5
|
|
; GFX11-FAKE16-NEXT: .LBB57_3:
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB57_2
|
|
; GFX11-FAKE16-NEXT: .LBB57_4:
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v23, s79 :: v_dual_mov_b32 v24, s89
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v25, s88 :: v_dual_mov_b32 v26, s78
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v27, s76 :: v_dual_mov_b32 v18, s74
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v19, s77 :: v_dual_mov_b32 v20, s75
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v21, s73 :: v_dual_mov_b32 v22, s72
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s29 :: v_dual_mov_b32 v14, s28
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v16, s26
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s25 :: v_dual_mov_b32 v8, s24
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v10, s22
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v12, s20
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, s19 :: v_dual_mov_b32 v4, s18
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s16
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v7, s3 :: v_dual_mov_b32 v0, s2
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s0
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v28, s59 :: v_dual_mov_b32 v29, s63
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v30, s62 :: v_dual_mov_b32 v31, s60
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v32, s57 :: v_dual_mov_b32 v33, s47
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v34, s61 :: v_dual_mov_b32 v35, s58
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, s56 :: v_dual_mov_b32 v37, s46
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v38, s45 :: v_dual_mov_b32 v39, s44
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v48, s43 :: v_dual_mov_b32 v49, s42
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v50, s41 :: v_dual_mov_b32 v51, s40
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v52, s15 :: v_dual_mov_b32 v53, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v54, s13 :: v_dual_mov_b32 v55, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v64, s11 :: v_dual_mov_b32 v65, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v66, s9 :: v_dual_mov_b32 v67, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v68, s8 :: v_dual_mov_b32 v69, s7
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v70, s4 :: v_dual_mov_b32 v71, s5
|
|
; GFX11-FAKE16-NEXT: .LBB57_5: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v80, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v69, 16, v80
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v69, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v70, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v68, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v67, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v65, 16, v69
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v64, 16, v70
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v65, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v55, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v54, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v52, 16, v64
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v51, 16, v65
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v13
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v50, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v49, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v39, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v38, 16, v52
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v19
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v23
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v31, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v29, 16, v33
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v28, 16, v34
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <56 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <56 x i16> %a1 to <56 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x i16> %a to <56 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x half> %phi
|
|
}
|
|
|
|
define <56 x i16> @bitcast_v56f16_to_v56i16(<56 x half> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v56f16_to_v56i16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v26
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v15
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v13
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v7
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v5
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v3
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v0
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; SI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB58_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v47
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v56
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v46
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v45
|
|
; SI-NEXT: v_add_f32_e32 v47, 0x38000000, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v47
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v44
|
|
; SI-NEXT: v_add_f32_e32 v45, 0x38000000, v45
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v47
|
|
; SI-NEXT: v_add_f32_e32 v47, 0x38000000, v56
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v47
|
|
; SI-NEXT: v_add_f32_e32 v44, 0x38000000, v44
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v56, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v47
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v41
|
|
; SI-NEXT: v_lshlrev_b32_e32 v43, 16, v44
|
|
; SI-NEXT: v_add_f32_e32 v42, 0x38000000, v42
|
|
; SI-NEXT: v_add_f32_e32 v44, 0x38000000, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v44
|
|
; SI-NEXT: v_add_f32_e32 v41, 0x38000000, v41
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v42
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v41
|
|
; SI-NEXT: v_lshlrev_b32_e32 v42, 16, v44
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v40
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v55
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v41
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v54
|
|
; SI-NEXT: v_add_f32_e32 v41, 0x38000000, v44
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v41
|
|
; SI-NEXT: v_add_f32_e32 v55, 0x38000000, v55
|
|
; SI-NEXT: v_add_f32_e32 v54, 0x38000000, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v55
|
|
; SI-NEXT: v_lshlrev_b32_e32 v55, 16, v41
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v53
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v30
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v54
|
|
; SI-NEXT: v_add_f32_e32 v54, 0x38000000, v41
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v52
|
|
; SI-NEXT: v_add_f32_e32 v29, 0x38000000, v29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_add_f32_e32 v27, 0x38000000, v27
|
|
; SI-NEXT: v_add_f32_e32 v30, 0x38000000, v30
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v34
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v30, v30
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_add_f32_e32 v31, 0x38000000, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, v35
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_add_f32_e32 v32, 0x38000000, v32
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v36
|
|
; SI-NEXT: v_add_f32_e32 v52, 0x38000000, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_add_f32_e32 v33, 0x38000000, v33
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v37
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v52
|
|
; SI-NEXT: v_lshlrev_b32_e32 v52, 16, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v33
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_add_f32_e32 v34, 0x38000000, v34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v38
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, v34
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_add_f32_e32 v35, 0x38000000, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v39
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, v35
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_add_f32_e32 v36, 0x38000000, v36
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v48
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v36
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_add_f32_e32 v37, 0x38000000, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v49
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v37
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_add_f32_e32 v38, 0x38000000, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v51
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v34
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v38
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_add_f32_e32 v39, 0x38000000, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v35
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v39
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_add_f32_e32 v48, 0x38000000, v48
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v36
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v48
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_add_f32_e32 v49, 0x38000000, v49
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v37
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v49
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_add_f32_e32 v50, 0x38000000, v50
|
|
; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28
|
|
; SI-NEXT: v_add_f32_e32 v51, 0x38000000, v51
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v38
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v50
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, v51
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v39
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v56
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v47
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v44
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v41
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v54
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v28
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v46
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v45
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v57
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v43
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v42
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v58
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v40
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v55
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v59
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v53
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v52
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v60
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v51
|
|
; SI-NEXT: v_alignbit_b32 v56, v1, v28, 16
|
|
; SI-NEXT: v_alignbit_b32 v47, v3, v46, 16
|
|
; SI-NEXT: v_alignbit_b32 v46, v5, v45, 16
|
|
; SI-NEXT: v_alignbit_b32 v45, v7, v57, 16
|
|
; SI-NEXT: v_alignbit_b32 v44, v9, v43, 16
|
|
; SI-NEXT: v_alignbit_b32 v43, v11, v42, 16
|
|
; SI-NEXT: v_alignbit_b32 v42, v13, v58, 16
|
|
; SI-NEXT: v_alignbit_b32 v41, v15, v40, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v17, v55, 16
|
|
; SI-NEXT: v_alignbit_b32 v55, v19, v59, 16
|
|
; SI-NEXT: v_alignbit_b32 v54, v21, v53, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v23, v52, 16
|
|
; SI-NEXT: v_alignbit_b32 v52, v25, v60, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v27, v51, 16
|
|
; SI-NEXT: .LBB58_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v56
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v28
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v28
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v28
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v28
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v28
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v28
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v28
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v28
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v28
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v28
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v28
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v28
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v28
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v28
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v41
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v28
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v28
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v40
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v28
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v28
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v28
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v28
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v28
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v28
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v28
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v31
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v28
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v28
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v30
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v28
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v29
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56f16_to_v56i16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v27
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v26
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v25
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v40, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v41, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v42, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; VI-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; VI-NEXT: s_cbranch_execz .LBB58_2
|
|
; VI-NEXT: ; %bb.1: ; %cmp.true
|
|
; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
|
|
; VI-NEXT: v_add_f16_e32 v56, 0x200, v56
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
|
|
; VI-NEXT: v_add_f16_e32 v47, 0x200, v47
|
|
; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
|
|
; VI-NEXT: v_add_f16_e32 v46, 0x200, v46
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
|
|
; VI-NEXT: v_add_f16_e32 v45, 0x200, v45
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
|
|
; VI-NEXT: v_add_f16_e32 v44, 0x200, v44
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
|
|
; VI-NEXT: v_add_f16_e32 v43, 0x200, v43
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
|
|
; VI-NEXT: v_add_f16_e32 v42, 0x200, v42
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
|
|
; VI-NEXT: v_add_f16_e32 v41, 0x200, v41
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
|
|
; VI-NEXT: v_add_f16_e32 v40, 0x200, v40
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
|
|
; VI-NEXT: v_add_f16_e32 v55, 0x200, v55
|
|
; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
|
|
; VI-NEXT: v_add_f16_e32 v54, 0x200, v54
|
|
; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
|
|
; VI-NEXT: v_add_f16_e32 v53, 0x200, v53
|
|
; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
|
|
; VI-NEXT: v_add_f16_e32 v52, 0x200, v52
|
|
; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
|
|
; VI-NEXT: v_add_f16_e32 v51, 0x200, v51
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
|
|
; VI-NEXT: v_add_f16_e32 v50, 0x200, v50
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
|
|
; VI-NEXT: v_add_f16_e32 v49, 0x200, v49
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
|
|
; VI-NEXT: v_add_f16_e32 v48, 0x200, v48
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v17
|
|
; VI-NEXT: v_add_f16_e32 v39, 0x200, v39
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v18
|
|
; VI-NEXT: v_add_f16_e32 v38, 0x200, v38
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v19
|
|
; VI-NEXT: v_add_f16_e32 v37, 0x200, v37
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v20
|
|
; VI-NEXT: v_add_f16_e32 v36, 0x200, v36
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v21
|
|
; VI-NEXT: v_add_f16_e32 v35, 0x200, v35
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v22
|
|
; VI-NEXT: v_add_f16_e32 v34, 0x200, v34
|
|
; VI-NEXT: v_add_f16_e32 v23, 0x200, v23
|
|
; VI-NEXT: v_add_f16_e32 v33, 0x200, v33
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v24
|
|
; VI-NEXT: v_add_f16_e32 v32, 0x200, v32
|
|
; VI-NEXT: v_add_f16_e32 v25, 0x200, v25
|
|
; VI-NEXT: v_add_f16_e32 v31, 0x200, v31
|
|
; VI-NEXT: v_add_f16_e32 v26, 0x200, v26
|
|
; VI-NEXT: v_add_f16_e32 v30, 0x200, v30
|
|
; VI-NEXT: v_add_f16_e32 v27, 0x200, v27
|
|
; VI-NEXT: v_add_f16_e32 v29, 0x200, v29
|
|
; VI-NEXT: .LBB58_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v56
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v47
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v46
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v45
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v44
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v43
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v42
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v41
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v40
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v55
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v54
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v53
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v52
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v51
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v50
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v49
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v48
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v39
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v38
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v37
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v36
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v35
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v34
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v33
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v32
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v31
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v30
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v29
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56f16_to_v56i16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v56, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v0
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v28
|
|
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
|
|
; GFX9-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
|
|
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
|
|
; GFX9-NEXT: s_cbranch_execz .LBB58_2
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v27, v56, v27, s6
|
|
; GFX9-NEXT: s_movk_i32 s7, 0x200
|
|
; GFX9-NEXT: v_perm_b32 v26, v47, v26, s6
|
|
; GFX9-NEXT: v_perm_b32 v25, v46, v25, s6
|
|
; GFX9-NEXT: v_perm_b32 v24, v45, v24, s6
|
|
; GFX9-NEXT: v_perm_b32 v23, v44, v23, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v43, v22, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v42, v21, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v41, v20, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v40, v19, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v55, v18, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v54, v17, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v53, v16, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v52, v15, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v51, v14, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v49, v12, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v48, v11, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v39, v10, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v37, v8, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v36, v7, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v35, v6, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v34, v5, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v4, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v32, v3, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v31, v2, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v30, v1, s6
|
|
; GFX9-NEXT: v_perm_b32 v0, v29, v0, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v27, v27, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v26, v26, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v25, v25, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v24, v24, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, v23, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, v22, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, v21, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, v20, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, v19, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, v18, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, v17, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, v16, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, v15, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, v14, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, v13, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, v12, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, v5, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, v4, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, v3, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, v2, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, v1, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v0, v0, s7 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v27
|
|
; GFX9-NEXT: .LBB58_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v19, v40, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v41, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v42, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v43, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v44, v23, s4
|
|
; GFX9-NEXT: v_perm_b32 v24, v45, v24, s4
|
|
; GFX9-NEXT: v_perm_b32 v25, v46, v25, s4
|
|
; GFX9-NEXT: v_perm_b32 v26, v47, v26, s4
|
|
; GFX9-NEXT: v_perm_b32 v27, v56, v27, s4
|
|
; GFX9-NEXT: buffer_load_dword v56, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_perm_b32 v0, v29, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v30, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v31, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v32, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v33, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v34, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v35, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v36, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v37, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v39, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v48, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v49, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v50, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v51, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v52, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v53, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v54, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v55, v18, s4
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v56i16:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB58_2
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: .LBB58_2: ; %end
|
|
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56f16_to_v56i16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v0
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-FAKE16-NEXT: v_cmpx_ne_u32_e32 0, v28
|
|
; GFX11-FAKE16-NEXT: s_xor_b32 s0, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
|
|
; GFX11-FAKE16-NEXT: s_and_not1_saveexec_b32 s0, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_execz .LBB58_2
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v80, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v71, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v70, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v69, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v68, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v67, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v66, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v65, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v64, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v55, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v54, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v53, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v52, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v51, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v49, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v48, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v39, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v37, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v36, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v35, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v29, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v30, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v31, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v32, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v33, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v27
|
|
; GFX11-FAKE16-NEXT: .LBB58_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v29, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v30, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v31, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v32, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v33, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v34, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v35, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v36, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v37, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v39, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v48, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v49, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v50, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v51, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v52, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v53, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v54, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v55, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v64, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v65, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v66, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v67, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v68, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v24, v69, v24, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v25, v70, v25, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v26, v71, v26, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v27, v80, v27, 0x5040100
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <56 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <56 x half> %a1 to <56 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x half> %a to <56 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x i16> %phi
|
|
}
|
|
|
|
define inreg <56 x i16> @bitcast_v56f16_to_v56i16_scalar(<56 x half> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v56f16_to_v56i16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_readfirstlane_b32 s44, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s60, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s42, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s62, v10
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v9
|
|
; SI-NEXT: v_readfirstlane_b32 s73, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s76, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s79, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s90, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s92, v0
|
|
; SI-NEXT: s_lshr_b32 s12, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s75, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s11, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s78, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s10, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s9, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s91, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s8, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s7, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s94, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s6, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s95, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s72, s44, 16
|
|
; SI-NEXT: s_lshr_b32 s46, s60, 16
|
|
; SI-NEXT: s_lshr_b32 s61, s42, 16
|
|
; SI-NEXT: s_lshr_b32 s57, s62, 16
|
|
; SI-NEXT: s_lshr_b32 s58, s41, 16
|
|
; SI-NEXT: s_lshr_b32 s59, s73, 16
|
|
; SI-NEXT: s_lshr_b32 s56, s40, 16
|
|
; SI-NEXT: s_lshr_b32 s63, s76, 16
|
|
; SI-NEXT: s_lshr_b32 s47, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s74, s79, 16
|
|
; SI-NEXT: s_lshr_b32 s45, s14, 16
|
|
; SI-NEXT: s_lshr_b32 s77, s90, 16
|
|
; SI-NEXT: s_lshr_b32 s43, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s88, s92, 16
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; SI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_cbranch_scc0 .LBB59_3
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB59_4
|
|
; SI-NEXT: .LBB59_2: ; %cmp.true
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, s95
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s94
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s18
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: v_or_b32_e32 v50, v1, v0
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s93
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v54, v3, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s20
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s91
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s22
|
|
; SI-NEXT: v_or_b32_e32 v51, v3, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s89
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v5
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_or_b32_e32 v48, v1, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s78
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s26
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_or_b32_e32 v49, v5, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s75
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_or_b32_e32 v34, v3, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s88
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s92
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_or_b32_e32 v56, v1, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s77
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s90
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_or_b32_e32 v35, v5, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s74
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s79
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_or_b32_e32 v33, v3, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s63
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s76
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_or_b32_e32 v32, v1, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s59
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s73
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_or_b32_e32 v31, v5, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s57
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s62
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v22
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s46
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s60
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s72
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s44
|
|
; SI-NEXT: v_or_b32_e32 v30, v5, v26
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s61
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v1
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v52
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s42
|
|
; SI-NEXT: v_or_b32_e32 v27, v1, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s58
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s41
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v57, v3
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v29
|
|
; SI-NEXT: v_or_b32_e32 v25, v5, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v57
|
|
; SI-NEXT: v_or_b32_e32 v23, v3, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s40
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v1
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s15
|
|
; SI-NEXT: v_or_b32_e32 v21, v1, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s45
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s14
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v3
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v19, v5, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s43
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v17, v3, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s13
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v62, v1
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v63, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v62
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s29
|
|
; SI-NEXT: v_or_b32_e32 v15, v1, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s27
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v3
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v63
|
|
; SI-NEXT: v_or_b32_e32 v13, v5, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v11, v3, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s9
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v60, v1
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v60
|
|
; SI-NEXT: v_or_b32_e32 v9, v1, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s8
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v36
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v58, v3
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v5, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v58
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, s17
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_or_b32_e32 v5, v3, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v59, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s6
|
|
; SI-NEXT: v_lshr_b64 v[42:43], v[4:5], 16
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v61, v1
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v59
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v28
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v61
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v28
|
|
; SI-NEXT: v_lshr_b64 v[46:47], v[0:1], 16
|
|
; SI-NEXT: v_mov_b32_e32 v4, v54
|
|
; SI-NEXT: v_lshr_b64 v[54:55], v[8:9], 16
|
|
; SI-NEXT: v_mov_b32_e32 v8, v48
|
|
; SI-NEXT: v_lshr_b64 v[47:48], v[10:11], 16
|
|
; SI-NEXT: v_lshr_b64 v[44:45], v[2:3], 16
|
|
; SI-NEXT: v_mov_b32_e32 v2, v50
|
|
; SI-NEXT: v_lshr_b64 v[40:41], v[6:7], 16
|
|
; SI-NEXT: v_mov_b32_e32 v6, v51
|
|
; SI-NEXT: v_mov_b32_e32 v10, v49
|
|
; SI-NEXT: v_lshr_b64 v[50:51], v[12:13], 16
|
|
; SI-NEXT: v_lshr_b64 v[48:49], v[14:15], 16
|
|
; SI-NEXT: v_mov_b32_e32 v51, v63
|
|
; SI-NEXT: v_mov_b32_e32 v12, v34
|
|
; SI-NEXT: v_mov_b32_e32 v49, v62
|
|
; SI-NEXT: v_mov_b32_e32 v14, v56
|
|
; SI-NEXT: v_lshr_b64 v[62:63], v[16:17], 16
|
|
; SI-NEXT: v_mov_b32_e32 v16, v35
|
|
; SI-NEXT: v_lshr_b64 v[55:56], v[18:19], 16
|
|
; SI-NEXT: v_mov_b32_e32 v18, v33
|
|
; SI-NEXT: v_lshr_b64 v[34:35], v[20:21], 16
|
|
; SI-NEXT: v_mov_b32_e32 v20, v32
|
|
; SI-NEXT: v_lshr_b64 v[32:33], v[22:23], 16
|
|
; SI-NEXT: v_mov_b32_e32 v33, v30
|
|
; SI-NEXT: v_mov_b32_e32 v22, v31
|
|
; SI-NEXT: v_lshr_b64 v[30:31], v[24:25], 16
|
|
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v31, v29
|
|
; SI-NEXT: v_lshr_b64 v[28:29], v[26:27], 16
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_branch .LBB59_5
|
|
; SI-NEXT: .LBB59_3:
|
|
; SI-NEXT: s_branch .LBB59_2
|
|
; SI-NEXT: .LBB59_4:
|
|
; SI-NEXT: v_mov_b32_e32 v52, s72
|
|
; SI-NEXT: v_mov_b32_e32 v31, s61
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: v_mov_b32_e32 v57, s58
|
|
; SI-NEXT: v_mov_b32_e32 v37, s56
|
|
; SI-NEXT: v_mov_b32_e32 v39, s47
|
|
; SI-NEXT: v_mov_b32_e32 v38, s45
|
|
; SI-NEXT: v_mov_b32_e32 v49, s43
|
|
; SI-NEXT: v_mov_b32_e32 v51, s12
|
|
; SI-NEXT: v_mov_b32_e32 v53, s11
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_mov_b32_e32 v60, s10
|
|
; SI-NEXT: v_mov_b32_e32 v36, s9
|
|
; SI-NEXT: v_mov_b32_e32 v58, s8
|
|
; SI-NEXT: v_mov_b32_e32 v59, s7
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v61, s6
|
|
; SI-NEXT: v_mov_b32_e32 v33, s60
|
|
; SI-NEXT: v_mov_b32_e32 v26, s62
|
|
; SI-NEXT: v_mov_b32_e32 v24, s73
|
|
; SI-NEXT: v_mov_b32_e32 v22, s76
|
|
; SI-NEXT: v_mov_b32_e32 v20, s79
|
|
; SI-NEXT: v_mov_b32_e32 v18, s90
|
|
; SI-NEXT: v_mov_b32_e32 v16, s92
|
|
; SI-NEXT: v_mov_b32_e32 v14, s28
|
|
; SI-NEXT: v_mov_b32_e32 v12, s26
|
|
; SI-NEXT: v_mov_b32_e32 v10, s24
|
|
; SI-NEXT: v_mov_b32_e32 v8, s22
|
|
; SI-NEXT: v_mov_b32_e32 v6, s20
|
|
; SI-NEXT: v_mov_b32_e32 v4, s18
|
|
; SI-NEXT: v_mov_b32_e32 v2, s16
|
|
; SI-NEXT: v_mov_b32_e32 v1, s17
|
|
; SI-NEXT: v_mov_b32_e32 v3, s19
|
|
; SI-NEXT: v_mov_b32_e32 v5, s21
|
|
; SI-NEXT: v_mov_b32_e32 v7, s23
|
|
; SI-NEXT: v_mov_b32_e32 v9, s25
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: v_mov_b32_e32 v15, s13
|
|
; SI-NEXT: v_mov_b32_e32 v17, s14
|
|
; SI-NEXT: v_mov_b32_e32 v19, s15
|
|
; SI-NEXT: v_mov_b32_e32 v21, s40
|
|
; SI-NEXT: v_mov_b32_e32 v23, s41
|
|
; SI-NEXT: v_mov_b32_e32 v25, s42
|
|
; SI-NEXT: v_mov_b32_e32 v27, s44
|
|
; SI-NEXT: v_mov_b32_e32 v46, s95
|
|
; SI-NEXT: v_mov_b32_e32 v44, s94
|
|
; SI-NEXT: v_mov_b32_e32 v42, s93
|
|
; SI-NEXT: v_mov_b32_e32 v40, s91
|
|
; SI-NEXT: v_mov_b32_e32 v54, s89
|
|
; SI-NEXT: v_mov_b32_e32 v47, s78
|
|
; SI-NEXT: v_mov_b32_e32 v50, s75
|
|
; SI-NEXT: v_mov_b32_e32 v48, s88
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_mov_b32_e32 v62, s77
|
|
; SI-NEXT: v_mov_b32_e32 v55, s74
|
|
; SI-NEXT: v_mov_b32_e32 v34, s63
|
|
; SI-NEXT: v_mov_b32_e32 v32, s59
|
|
; SI-NEXT: v_mov_b32_e32 v30, s57
|
|
; SI-NEXT: v_mov_b32_e32 v28, s46
|
|
; SI-NEXT: .LBB59_5: ; %end
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v46
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; SI-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v61
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v44
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; SI-NEXT: v_or_b32_e32 v2, v4, v2
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v59
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v42
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: v_or_b32_e32 v4, v6, v4
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v58
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v40
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_or_b32_e32 v6, v8, v6
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v54
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: v_or_b32_e32 v8, v10, v8
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v60
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v47
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_or_b32_e32 v10, v12, v10
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v14
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v16
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v62
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v18
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v20
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v22
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v57
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_and_b32_e32 v24, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v30
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v26
|
|
; SI-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v31
|
|
; SI-NEXT: v_or_b32_e32 v25, v25, v26
|
|
; SI-NEXT: v_and_b32_e32 v26, 0xffff, v33
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v26, v26, v28
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v56f16_to_v56i16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: v_writelane_b32 v56, s30, 0
|
|
; VI-NEXT: v_writelane_b32 v56, s31, 1
|
|
; VI-NEXT: v_writelane_b32 v56, s34, 2
|
|
; VI-NEXT: v_readfirstlane_b32 s44, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s46, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s56, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s58, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s60, v9
|
|
; VI-NEXT: v_readfirstlane_b32 s62, v8
|
|
; VI-NEXT: v_readfirstlane_b32 s72, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s74, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s76, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s78, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s88, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s90, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s30, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s34, v0
|
|
; VI-NEXT: v_writelane_b32 v56, s35, 3
|
|
; VI-NEXT: s_lshr_b32 s6, s29, 16
|
|
; VI-NEXT: s_lshr_b32 s7, s28, 16
|
|
; VI-NEXT: s_lshr_b32 s8, s27, 16
|
|
; VI-NEXT: s_lshr_b32 s9, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s10, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s11, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s12, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s13, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s14, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s15, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s40, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s41, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s42, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s43, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s44, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s46, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s56, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s58, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s60, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s62, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s72, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s74, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s76, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s78, 16
|
|
; VI-NEXT: s_lshr_b32 s89, s88, 16
|
|
; VI-NEXT: s_lshr_b32 s91, s90, 16
|
|
; VI-NEXT: s_lshr_b32 s31, s30, 16
|
|
; VI-NEXT: s_lshr_b32 s35, s34, 16
|
|
; VI-NEXT: v_readfirstlane_b32 s4, v14
|
|
; VI-NEXT: s_cmp_lg_u32 s4, 0
|
|
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; VI-NEXT: s_cbranch_scc0 .LBB59_3
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB59_4
|
|
; VI-NEXT: .LBB59_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v28, 0x200
|
|
; VI-NEXT: v_add_f16_e32 v0, s16, v28
|
|
; VI-NEXT: v_add_f16_e32 v47, s43, v28
|
|
; VI-NEXT: v_add_f16_e32 v1, s17, v28
|
|
; VI-NEXT: v_add_f16_e32 v46, s42, v28
|
|
; VI-NEXT: v_add_f16_e32 v2, s18, v28
|
|
; VI-NEXT: v_add_f16_e32 v45, s41, v28
|
|
; VI-NEXT: v_add_f16_e32 v3, s19, v28
|
|
; VI-NEXT: v_add_f16_e32 v44, s40, v28
|
|
; VI-NEXT: v_add_f16_e32 v4, s20, v28
|
|
; VI-NEXT: v_add_f16_e32 v43, s15, v28
|
|
; VI-NEXT: v_add_f16_e32 v5, s21, v28
|
|
; VI-NEXT: v_add_f16_e32 v42, s14, v28
|
|
; VI-NEXT: v_add_f16_e32 v6, s22, v28
|
|
; VI-NEXT: v_add_f16_e32 v41, s13, v28
|
|
; VI-NEXT: v_add_f16_e32 v7, s23, v28
|
|
; VI-NEXT: v_add_f16_e32 v40, s12, v28
|
|
; VI-NEXT: v_add_f16_e32 v8, s24, v28
|
|
; VI-NEXT: v_add_f16_e32 v55, s11, v28
|
|
; VI-NEXT: v_add_f16_e32 v9, s25, v28
|
|
; VI-NEXT: v_add_f16_e32 v54, s10, v28
|
|
; VI-NEXT: v_add_f16_e32 v10, s26, v28
|
|
; VI-NEXT: v_add_f16_e32 v53, s9, v28
|
|
; VI-NEXT: v_add_f16_e32 v11, s27, v28
|
|
; VI-NEXT: v_add_f16_e32 v52, s8, v28
|
|
; VI-NEXT: v_add_f16_e32 v12, s28, v28
|
|
; VI-NEXT: v_add_f16_e32 v51, s7, v28
|
|
; VI-NEXT: v_add_f16_e32 v13, s29, v28
|
|
; VI-NEXT: v_add_f16_e32 v50, s6, v28
|
|
; VI-NEXT: v_add_f16_e32 v14, s34, v28
|
|
; VI-NEXT: v_add_f16_e32 v49, s35, v28
|
|
; VI-NEXT: v_add_f16_e32 v15, s30, v28
|
|
; VI-NEXT: v_add_f16_e32 v48, s31, v28
|
|
; VI-NEXT: v_add_f16_e32 v16, s90, v28
|
|
; VI-NEXT: v_add_f16_e32 v39, s91, v28
|
|
; VI-NEXT: v_add_f16_e32 v17, s88, v28
|
|
; VI-NEXT: v_add_f16_e32 v38, s89, v28
|
|
; VI-NEXT: v_add_f16_e32 v18, s78, v28
|
|
; VI-NEXT: v_add_f16_e32 v37, s79, v28
|
|
; VI-NEXT: v_add_f16_e32 v19, s76, v28
|
|
; VI-NEXT: v_add_f16_e32 v36, s77, v28
|
|
; VI-NEXT: v_add_f16_e32 v20, s74, v28
|
|
; VI-NEXT: v_add_f16_e32 v35, s75, v28
|
|
; VI-NEXT: v_add_f16_e32 v21, s72, v28
|
|
; VI-NEXT: v_add_f16_e32 v34, s73, v28
|
|
; VI-NEXT: v_add_f16_e32 v22, s62, v28
|
|
; VI-NEXT: v_add_f16_e32 v33, s63, v28
|
|
; VI-NEXT: v_add_f16_e32 v23, s60, v28
|
|
; VI-NEXT: v_add_f16_e32 v32, s61, v28
|
|
; VI-NEXT: v_add_f16_e32 v24, s58, v28
|
|
; VI-NEXT: v_add_f16_e32 v31, s59, v28
|
|
; VI-NEXT: v_add_f16_e32 v25, s56, v28
|
|
; VI-NEXT: v_add_f16_e32 v30, s57, v28
|
|
; VI-NEXT: v_add_f16_e32 v26, s46, v28
|
|
; VI-NEXT: v_add_f16_e32 v29, s47, v28
|
|
; VI-NEXT: v_add_f16_e32 v27, s44, v28
|
|
; VI-NEXT: v_add_f16_e32 v28, s45, v28
|
|
; VI-NEXT: s_branch .LBB59_5
|
|
; VI-NEXT: .LBB59_3:
|
|
; VI-NEXT: s_branch .LBB59_2
|
|
; VI-NEXT: .LBB59_4:
|
|
; VI-NEXT: v_mov_b32_e32 v28, s45
|
|
; VI-NEXT: v_mov_b32_e32 v27, s44
|
|
; VI-NEXT: v_mov_b32_e32 v29, s47
|
|
; VI-NEXT: v_mov_b32_e32 v26, s46
|
|
; VI-NEXT: v_mov_b32_e32 v30, s57
|
|
; VI-NEXT: v_mov_b32_e32 v25, s56
|
|
; VI-NEXT: v_mov_b32_e32 v31, s59
|
|
; VI-NEXT: v_mov_b32_e32 v24, s58
|
|
; VI-NEXT: v_mov_b32_e32 v32, s61
|
|
; VI-NEXT: v_mov_b32_e32 v23, s60
|
|
; VI-NEXT: v_mov_b32_e32 v33, s63
|
|
; VI-NEXT: v_mov_b32_e32 v22, s62
|
|
; VI-NEXT: v_mov_b32_e32 v34, s73
|
|
; VI-NEXT: v_mov_b32_e32 v21, s72
|
|
; VI-NEXT: v_mov_b32_e32 v35, s75
|
|
; VI-NEXT: v_mov_b32_e32 v20, s74
|
|
; VI-NEXT: v_mov_b32_e32 v36, s77
|
|
; VI-NEXT: v_mov_b32_e32 v19, s76
|
|
; VI-NEXT: v_mov_b32_e32 v37, s79
|
|
; VI-NEXT: v_mov_b32_e32 v18, s78
|
|
; VI-NEXT: v_mov_b32_e32 v38, s89
|
|
; VI-NEXT: v_mov_b32_e32 v17, s88
|
|
; VI-NEXT: v_mov_b32_e32 v39, s91
|
|
; VI-NEXT: v_mov_b32_e32 v16, s90
|
|
; VI-NEXT: v_mov_b32_e32 v48, s31
|
|
; VI-NEXT: v_mov_b32_e32 v15, s30
|
|
; VI-NEXT: v_mov_b32_e32 v49, s35
|
|
; VI-NEXT: v_mov_b32_e32 v14, s34
|
|
; VI-NEXT: v_mov_b32_e32 v50, s6
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v51, s7
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v52, s8
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v53, s9
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v54, s10
|
|
; VI-NEXT: v_mov_b32_e32 v9, s25
|
|
; VI-NEXT: v_mov_b32_e32 v55, s11
|
|
; VI-NEXT: v_mov_b32_e32 v8, s24
|
|
; VI-NEXT: v_mov_b32_e32 v40, s12
|
|
; VI-NEXT: v_mov_b32_e32 v7, s23
|
|
; VI-NEXT: v_mov_b32_e32 v41, s13
|
|
; VI-NEXT: v_mov_b32_e32 v6, s22
|
|
; VI-NEXT: v_mov_b32_e32 v42, s14
|
|
; VI-NEXT: v_mov_b32_e32 v5, s21
|
|
; VI-NEXT: v_mov_b32_e32 v43, s15
|
|
; VI-NEXT: v_mov_b32_e32 v4, s20
|
|
; VI-NEXT: v_mov_b32_e32 v44, s40
|
|
; VI-NEXT: v_mov_b32_e32 v3, s19
|
|
; VI-NEXT: v_mov_b32_e32 v45, s41
|
|
; VI-NEXT: v_mov_b32_e32 v2, s18
|
|
; VI-NEXT: v_mov_b32_e32 v46, s42
|
|
; VI-NEXT: v_mov_b32_e32 v1, s17
|
|
; VI-NEXT: v_mov_b32_e32 v47, s43
|
|
; VI-NEXT: v_mov_b32_e32 v0, s16
|
|
; VI-NEXT: .LBB59_5: ; %end
|
|
; VI-NEXT: v_lshlrev_b32_e32 v47, 16, v47
|
|
; VI-NEXT: v_lshlrev_b32_e32 v46, 16, v46
|
|
; VI-NEXT: v_lshlrev_b32_e32 v45, 16, v45
|
|
; VI-NEXT: v_lshlrev_b32_e32 v44, 16, v44
|
|
; VI-NEXT: v_lshlrev_b32_e32 v43, 16, v43
|
|
; VI-NEXT: v_lshlrev_b32_e32 v42, 16, v42
|
|
; VI-NEXT: v_lshlrev_b32_e32 v41, 16, v41
|
|
; VI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; VI-NEXT: v_lshlrev_b32_e32 v55, 16, v55
|
|
; VI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; VI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; VI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; VI-NEXT: v_lshlrev_b32_e32 v51, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v49, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v48, 16, v48
|
|
; VI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; VI-NEXT: v_lshlrev_b32_e32 v38, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; VI-NEXT: v_lshlrev_b32_e32 v36, 16, v36
|
|
; VI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; VI-NEXT: v_lshlrev_b32_e32 v34, 16, v34
|
|
; VI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; VI-NEXT: v_lshlrev_b32_e32 v32, 16, v32
|
|
; VI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v28, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_readlane_b32 s35, v56, 3
|
|
; VI-NEXT: v_readlane_b32 s34, v56, 2
|
|
; VI-NEXT: v_readlane_b32 s31, v56, 1
|
|
; VI-NEXT: v_readlane_b32 s30, v56, 0
|
|
; VI-NEXT: s_or_saveexec_b64 s[4:5], -1
|
|
; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; VI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; VI-NEXT: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v56f16_to_v56i16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_readfirstlane_b32 s95, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s94, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s93, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s92, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s91, v9
|
|
; GFX9-NEXT: v_readfirstlane_b32 s90, v8
|
|
; GFX9-NEXT: v_readfirstlane_b32 s89, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s88, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s79, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s78, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s77, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s76, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s75, v1
|
|
; GFX9-NEXT: v_readfirstlane_b32 s74, v0
|
|
; GFX9-NEXT: s_lshr_b32 s43, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s40, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s15, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s14, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s13, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s12, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s11, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s10, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s9, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s8, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s7, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s6, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s95, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s94, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s93, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s92, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s91, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s90, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s89, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s88, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s79, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s78, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s77, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s76, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s75, 16
|
|
; GFX9-NEXT: s_lshr_b32 s44, s74, 16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s4, v14
|
|
; GFX9-NEXT: s_cmp_lg_u32 s4, 0
|
|
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB59_3
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB59_4
|
|
; GFX9-NEXT: .LBB59_2: ; %cmp.true
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s95, s73
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, 0x200
|
|
; GFX9-NEXT: v_pk_add_f16 v27, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s94, s72
|
|
; GFX9-NEXT: v_pk_add_f16 v26, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s93, s63
|
|
; GFX9-NEXT: v_pk_add_f16 v25, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s92, s62
|
|
; GFX9-NEXT: v_pk_add_f16 v24, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s91, s61
|
|
; GFX9-NEXT: v_pk_add_f16 v23, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s90, s60
|
|
; GFX9-NEXT: v_pk_add_f16 v22, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s89, s59
|
|
; GFX9-NEXT: v_pk_add_f16 v21, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s88, s58
|
|
; GFX9-NEXT: v_pk_add_f16 v20, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s79, s57
|
|
; GFX9-NEXT: v_pk_add_f16 v19, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s78, s56
|
|
; GFX9-NEXT: v_pk_add_f16 v18, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s77, s47
|
|
; GFX9-NEXT: v_pk_add_f16 v17, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s76, s46
|
|
; GFX9-NEXT: v_pk_add_f16 v16, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s75, s45
|
|
; GFX9-NEXT: v_pk_add_f16 v15, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s74, s44
|
|
; GFX9-NEXT: v_pk_add_f16 v14, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s29, s43
|
|
; GFX9-NEXT: v_pk_add_f16 v13, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s28, s42
|
|
; GFX9-NEXT: v_pk_add_f16 v12, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s27, s41
|
|
; GFX9-NEXT: v_pk_add_f16 v11, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s26, s40
|
|
; GFX9-NEXT: v_pk_add_f16 v10, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s25, s15
|
|
; GFX9-NEXT: v_pk_add_f16 v9, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s24, s14
|
|
; GFX9-NEXT: v_pk_add_f16 v8, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s23, s13
|
|
; GFX9-NEXT: v_pk_add_f16 v7, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s22, s12
|
|
; GFX9-NEXT: v_pk_add_f16 v6, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s21, s11
|
|
; GFX9-NEXT: v_pk_add_f16 v5, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s20, s10
|
|
; GFX9-NEXT: v_pk_add_f16 v4, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s19, s9
|
|
; GFX9-NEXT: v_pk_add_f16 v3, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s18, s8
|
|
; GFX9-NEXT: v_pk_add_f16 v2, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s17, s7
|
|
; GFX9-NEXT: v_pk_add_f16 v1, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v0, s4, v0 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v47, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v46, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v45, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v44, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX9-NEXT: s_branch .LBB59_5
|
|
; GFX9-NEXT: .LBB59_3:
|
|
; GFX9-NEXT: s_branch .LBB59_2
|
|
; GFX9-NEXT: .LBB59_4:
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s95
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s94
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s93
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s92
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s91
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s90
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s89
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s88
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s79
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s78
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s77
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s76
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s75
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s74
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s73
|
|
; GFX9-NEXT: v_mov_b32_e32 v29, s72
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s63
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s62
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, s61
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, s60
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, s59
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, s58
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, s57
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, s56
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, s47
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, s46
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, s45
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, s44
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, s6
|
|
; GFX9-NEXT: .LBB59_5: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX9-NEXT: v_lshl_or_b32 v0, v47, 16, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v1, v46, 16, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v2, v45, 16, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v3, v44, 16, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v4, v43, 16, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v5, v42, 16, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v6, v41, 16, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v7, v40, 16, v7
|
|
; GFX9-NEXT: buffer_load_dword v47, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX9-NEXT: v_lshl_or_b32 v8, v55, 16, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v9, v54, 16, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v53, 16, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v52, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v51, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v50, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v49, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v48, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v39, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v38, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v37, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v36, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v35, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v34, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v33, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v32, 16, v23
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v31, 16, v24
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v30, 16, v25
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v29, 16, v26
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v28, 16, v27
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v56i16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s79, v9
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s89, v8
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s88, v7
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s78, v6
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s76, v5
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s74, v4
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s77, v3
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s75, v2
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s73, v1
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s72, v0
|
|
; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s90, v10
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s59, s79, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s63, s89, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s62, s88, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s60, s78, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s57, s76, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s47, s74, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s61, s77, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s58, s75, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s56, s73, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s72, 16
|
|
; GFX11-TRUE16-NEXT: s_cmp_lg_u32 s90, 0
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s90, 0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB59_3
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s90
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB59_4
|
|
; GFX11-TRUE16-NEXT: .LBB59_2: ; %cmp.true
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s59, s79, s59
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s57, s76, s57
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, s59 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s59, s88, s62
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s47, s74, s47
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s63, s89, s63
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, s59 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s59, s78, s60
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, s57 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s57, s77, s61
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, s47 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s47, s75, s58
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s56, s73, s56
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s46, s72, s46
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s29, s29, s45
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s28, s28, s44
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s27, s27, s43
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s26, s26, s42
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s25, s25, s41
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s24, s24, s40
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s23, s15
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s22, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s21, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s20, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s19, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s18, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s17, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s4
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, s63 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, s59 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, s57 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, s47 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, s56 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, s46 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, s29 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, s28 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s26 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, s25 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, s24 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 16, v0
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v67, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v65, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v26
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v27
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB59_5
|
|
; GFX11-TRUE16-NEXT: .LBB59_3:
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB59_2
|
|
; GFX11-TRUE16-NEXT: .LBB59_4:
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, s79 :: v_dual_mov_b32 v26, s89
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, s88 :: v_dual_mov_b32 v24, s78
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, s76 :: v_dual_mov_b32 v22, s74
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, s77 :: v_dual_mov_b32 v20, s75
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, s73 :: v_dual_mov_b32 v18, s72
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, s29 :: v_dual_mov_b32 v16, s28
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v14, s26
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v13, s25 :: v_dual_mov_b32 v12, s24
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v10, s22
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, s21 :: v_dual_mov_b32 v8, s20
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v6, s18
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v4, s16
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v0, s0
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, s59 :: v_dual_mov_b32 v29, s63
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, s62 :: v_dual_mov_b32 v31, s60
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v32, s57 :: v_dual_mov_b32 v33, s47
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v34, s61 :: v_dual_mov_b32 v35, s58
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v36, s56 :: v_dual_mov_b32 v37, s46
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v38, s45 :: v_dual_mov_b32 v39, s44
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v48, s43 :: v_dual_mov_b32 v49, s42
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v50, s41 :: v_dual_mov_b32 v51, s40
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v52, s15 :: v_dual_mov_b32 v53, s14
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s13 :: v_dual_mov_b32 v55, s12
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, s11 :: v_dual_mov_b32 v65, s10
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v66, s9 :: v_dual_mov_b32 v67, s6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v68, s8 :: v_dual_mov_b32 v69, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v70, s4 :: v_dual_mov_b32 v71, s5
|
|
; GFX11-TRUE16-NEXT: .LBB59_5: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v71.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v70.l
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v69.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v68.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v67.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v66.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v65.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v64.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v31.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v28.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v56f16_to_v56i16_scalar:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s79, v9
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s89, v8
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s88, v7
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s78, v6
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s76, v5
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s74, v4
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s77, v3
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s75, v2
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s73, v1
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s72, v0
|
|
; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s90, v10
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s40, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s59, s79, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s63, s89, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s62, s88, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s60, s78, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s57, s76, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s47, s74, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s61, s77, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s58, s75, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s56, s73, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s72, 16
|
|
; GFX11-FAKE16-NEXT: s_cmp_lg_u32 s90, 0
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s90, 0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB59_3
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s90
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB59_4
|
|
; GFX11-FAKE16-NEXT: .LBB59_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s59, s79, s59
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s57, s76, s57
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v23, 0x200, s59 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s59, s88, s62
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s47, s74, s47
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s63, s89, s63
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v25, 0x200, s59 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s59, s78, s60
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v27, 0x200, s57 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s57, s77, s61
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v18, 0x200, s47 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s47, s75, s58
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s56, s73, s56
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s46, s72, s46
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s29, s29, s45
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s28, s28, s44
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s27, s27, s43
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s26, s26, s42
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s25, s25, s41
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s24, s24, s40
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s23, s15
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s22, s14
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s21, s13
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s20, s12
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s19, s11
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s18, s10
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s17, s9
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s8
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s0, s5
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s4
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v24, 0x200, s63 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v26, 0x200, s59 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v19, 0x200, s57 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v20, 0x200, s47 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v21, 0x200, s56 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v22, 0x200, s46 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, s29 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, s28 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, s26 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, s25 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, s24 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v23
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB59_5
|
|
; GFX11-FAKE16-NEXT: .LBB59_3:
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB59_2
|
|
; GFX11-FAKE16-NEXT: .LBB59_4:
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v23, s79 :: v_dual_mov_b32 v24, s89
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v25, s88 :: v_dual_mov_b32 v26, s78
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v27, s76 :: v_dual_mov_b32 v18, s74
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v19, s77 :: v_dual_mov_b32 v20, s75
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v21, s73 :: v_dual_mov_b32 v22, s72
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s29 :: v_dual_mov_b32 v14, s28
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v16, s26
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s25 :: v_dual_mov_b32 v8, s24
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v10, s22
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v12, s20
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, s19 :: v_dual_mov_b32 v4, s18
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s16
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v7, s3 :: v_dual_mov_b32 v0, s2
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s0
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v28, s59 :: v_dual_mov_b32 v29, s63
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v30, s62 :: v_dual_mov_b32 v31, s60
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v32, s57 :: v_dual_mov_b32 v33, s47
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v34, s61 :: v_dual_mov_b32 v35, s58
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, s56 :: v_dual_mov_b32 v37, s46
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v38, s45 :: v_dual_mov_b32 v39, s44
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v48, s43 :: v_dual_mov_b32 v49, s42
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v50, s41 :: v_dual_mov_b32 v51, s40
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v52, s15 :: v_dual_mov_b32 v53, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v54, s13 :: v_dual_mov_b32 v55, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v64, s11 :: v_dual_mov_b32 v65, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v66, s9 :: v_dual_mov_b32 v67, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v68, s8 :: v_dual_mov_b32 v69, s7
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v70, s4 :: v_dual_mov_b32 v71, s5
|
|
; GFX11-FAKE16-NEXT: .LBB59_5: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v80, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v71, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v70, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v69, 16, v80
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v69, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v70, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v68, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v67, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v65, 16, v69
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v64, 16, v70
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v65, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v55, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v54, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v52, 16, v64
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v51, 16, v65
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v13
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v50, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v49, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v39, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v38, 16, v52
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v19
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v33, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v23
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v66, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v53, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v48, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v31, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v30, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v29, 16, v33
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v28, 16, v34
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <56 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <56 x half> %a1 to <56 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <56 x half> %a to <56 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <56 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <56 x i16> %phi
|
|
}
|