b1c4b55118
Attempt to only define used subregisters when creating IMPLICIT_DEF fix ups for live interval subranges. This avoids the appearance at the MIR level of entire (wide) registers becoming live rather than relying only on transient LiveIntervals dead definitions for unused subregisters.
41590 lines
1.9 MiB
41590 lines
1.9 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 <24 x float> @bitcast_v24i32_to_v24f32(<24 x i32> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v24i32_to_v24f32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24i32_to_v24f32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24i32_to_v24f32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24i32_to_v24f32:
|
|
; 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, v24
|
|
; 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 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 <24 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <24 x i32> %a1 to <24 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x i32> %a to <24 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x float> %phi
|
|
}
|
|
|
|
define inreg <24 x float> @bitcast_v24i32_to_v24f32_scalar(<24 x i32> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v24i32_to_v24f32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB1_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB1_4:
|
|
; SI-NEXT: s_branch .LBB1_2
|
|
;
|
|
; VI-LABEL: bitcast_v24i32_to_v24f32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB1_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB1_4:
|
|
; VI-NEXT: s_branch .LBB1_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v24i32_to_v24f32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB1_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB1_4:
|
|
; GFX9-NEXT: s_branch .LBB1_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v24i32_to_v24f32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB1_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccz .LBB1_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB1_3:
|
|
; GFX11-NEXT: .LBB1_4: ; %cmp.true
|
|
; 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: 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 <24 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <24 x i32> %a1 to <24 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x i32> %a to <24 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x float> %phi
|
|
}
|
|
|
|
define <24 x i32> @bitcast_v24f32_to_v24i32(<24 x float> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v24f32_to_v24i32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24f32_to_v24i32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24f32_to_v24i32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24f32_to_v24i32:
|
|
; 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, v24
|
|
; 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 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 <24 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <24 x float> %a1 to <24 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x float> %a to <24 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x i32> %phi
|
|
}
|
|
|
|
define inreg <24 x i32> @bitcast_v24f32_to_v24i32_scalar(<24 x float> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v24f32_to_v24i32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: s_cbranch_scc0 .LBB3_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB3_3
|
|
; SI-NEXT: .LBB3_2: ; %cmp.true
|
|
; 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: .LBB3_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB3_4:
|
|
; SI-NEXT: s_branch .LBB3_2
|
|
;
|
|
; VI-LABEL: bitcast_v24f32_to_v24i32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: s_cbranch_scc0 .LBB3_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB3_3
|
|
; VI-NEXT: .LBB3_2: ; %cmp.true
|
|
; 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: .LBB3_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB3_4:
|
|
; VI-NEXT: s_branch .LBB3_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v24f32_to_v24i32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: s_cbranch_scc0 .LBB3_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB3_3
|
|
; GFX9-NEXT: .LBB3_2: ; %cmp.true
|
|
; 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: .LBB3_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB3_4:
|
|
; GFX9-NEXT: s_branch .LBB3_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v24f32_to_v24i32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; 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_vccz .LBB3_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB3_3:
|
|
; GFX11-NEXT: .LBB3_4: ; %cmp.true
|
|
; 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: 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 <24 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <24 x float> %a1 to <24 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x float> %a to <24 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x i32> %phi
|
|
}
|
|
|
|
define <12 x i64> @bitcast_v24i32_to_v12i64(<24 x i32> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v24i32_to_v12i64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24i32_to_v12i64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24i32_to_v12i64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24i32_to_v12i64:
|
|
; 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, v24
|
|
; 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 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 <24 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <24 x i32> %a1 to <12 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x i32> %a to <12 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x i64> %phi
|
|
}
|
|
|
|
define inreg <12 x i64> @bitcast_v24i32_to_v12i64_scalar(<24 x i32> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v24i32_to_v12i64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB5_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB5_4:
|
|
; SI-NEXT: s_branch .LBB5_2
|
|
;
|
|
; VI-LABEL: bitcast_v24i32_to_v12i64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB5_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB5_4:
|
|
; VI-NEXT: s_branch .LBB5_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v24i32_to_v12i64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB5_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB5_4:
|
|
; GFX9-NEXT: s_branch .LBB5_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v24i32_to_v12i64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB5_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccz .LBB5_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB5_3:
|
|
; GFX11-NEXT: .LBB5_4: ; %cmp.true
|
|
; 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: 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 <24 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <24 x i32> %a1 to <12 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x i32> %a to <12 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x i64> %phi
|
|
}
|
|
|
|
define <24 x i32> @bitcast_v12i64_to_v24i32(<12 x i64> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v12i64_to_v24i32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v12i64_to_v24i32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v12i64_to_v24i32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v12i64_to_v24i32:
|
|
; 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, v24
|
|
; 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 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 <12 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <12 x i64> %a1 to <24 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x i64> %a to <24 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x i32> %phi
|
|
}
|
|
|
|
define inreg <24 x i32> @bitcast_v12i64_to_v24i32_scalar(<12 x i64> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v12i64_to_v24i32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB7_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB7_4:
|
|
; SI-NEXT: s_branch .LBB7_2
|
|
;
|
|
; VI-LABEL: bitcast_v12i64_to_v24i32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB7_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB7_4:
|
|
; VI-NEXT: s_branch .LBB7_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v12i64_to_v24i32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB7_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB7_4:
|
|
; GFX9-NEXT: s_branch .LBB7_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v12i64_to_v24i32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB7_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccz .LBB7_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB7_3:
|
|
; GFX11-NEXT: .LBB7_4: ; %cmp.true
|
|
; 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: 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 <12 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <12 x i64> %a1 to <24 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x i64> %a to <24 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x i32> %phi
|
|
}
|
|
|
|
define <12 x double> @bitcast_v24i32_to_v12f64(<24 x i32> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v24i32_to_v12f64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24i32_to_v12f64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24i32_to_v12f64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24i32_to_v12f64:
|
|
; 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, v24
|
|
; 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 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 <24 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <24 x i32> %a1 to <12 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x i32> %a to <12 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x double> %phi
|
|
}
|
|
|
|
define inreg <12 x double> @bitcast_v24i32_to_v12f64_scalar(<24 x i32> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v24i32_to_v12f64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB9_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB9_4:
|
|
; SI-NEXT: s_branch .LBB9_2
|
|
;
|
|
; VI-LABEL: bitcast_v24i32_to_v12f64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB9_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB9_4:
|
|
; VI-NEXT: s_branch .LBB9_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v24i32_to_v12f64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB9_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB9_4:
|
|
; GFX9-NEXT: s_branch .LBB9_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v24i32_to_v12f64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB9_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccz .LBB9_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB9_3:
|
|
; GFX11-NEXT: .LBB9_4: ; %cmp.true
|
|
; 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: 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 <24 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <24 x i32> %a1 to <12 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x i32> %a to <12 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x double> %phi
|
|
}
|
|
|
|
define <24 x i32> @bitcast_v12f64_to_v24i32(<12 x double> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v12f64_to_v24i32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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[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_v12f64_to_v24i32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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[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_v12f64_to_v24i32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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[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_v12f64_to_v24i32:
|
|
; 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, v24
|
|
; 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[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 <12 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <12 x double> %a1 to <24 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x double> %a to <24 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x i32> %phi
|
|
}
|
|
|
|
define inreg <24 x i32> @bitcast_v12f64_to_v24i32_scalar(<12 x double> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v12f64_to_v24i32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: s_cbranch_scc0 .LBB11_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB11_3
|
|
; SI-NEXT: .LBB11_2: ; %cmp.true
|
|
; 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: .LBB11_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB11_4:
|
|
; SI-NEXT: s_branch .LBB11_2
|
|
;
|
|
; VI-LABEL: bitcast_v12f64_to_v24i32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: s_cbranch_scc0 .LBB11_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB11_3
|
|
; VI-NEXT: .LBB11_2: ; %cmp.true
|
|
; 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: .LBB11_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB11_4:
|
|
; VI-NEXT: s_branch .LBB11_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v12f64_to_v24i32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB11_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB11_3
|
|
; GFX9-NEXT: .LBB11_2: ; %cmp.true
|
|
; 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: .LBB11_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB11_4:
|
|
; GFX9-NEXT: s_branch .LBB11_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v12f64_to_v24i32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; 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_vccz .LBB11_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB11_3:
|
|
; GFX11-NEXT: .LBB11_4: ; %cmp.true
|
|
; 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: 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 <12 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <12 x double> %a1 to <24 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x double> %a to <24 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x i32> %phi
|
|
}
|
|
|
|
define <48 x i16> @bitcast_v24i32_to_v48i16(<24 x i32> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v24i32_to_v48i16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v25
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; 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 v25, v24, v23, 16
|
|
; SI-NEXT: v_alignbit_b32 v26, v22, v21, 16
|
|
; SI-NEXT: v_alignbit_b32 v27, v20, v19, 16
|
|
; SI-NEXT: v_alignbit_b32 v28, v18, v17, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v16, v15, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v14, v13, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v12, v11, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v10, v9, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v8, v7, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v6, v5, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v4, v3, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v2, v1, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v2
|
|
; 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 v2, vcc, 3, v2
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
|
|
; 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 v6, vcc, 3, v6
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; 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 v10, vcc, 3, v10
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; 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 v14, vcc, 3, v14
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; 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 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; 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 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_alignbit_b32 v25, v24, v23, 16
|
|
; SI-NEXT: v_alignbit_b32 v26, v22, v21, 16
|
|
; SI-NEXT: v_alignbit_b32 v27, v20, v19, 16
|
|
; SI-NEXT: v_alignbit_b32 v28, v18, v17, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v16, v15, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v14, v13, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v12, v11, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v10, v9, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v8, v7, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v6, v5, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v4, v3, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v2, v1, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v2
|
|
; SI-NEXT: .LBB12_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v53
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 4, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 8, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 12, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 16, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 20, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 24, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 28, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 32, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 36, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 40, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v31
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 52, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v29
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 56, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 60, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 64, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v27
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x4c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x50, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x54, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x58, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v30
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v24i32_to_v48i16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; VI-NEXT: .LBB12_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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_lshlrev_b32_e32 v27, 16, v27
|
|
; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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_or_b32_sdwa v20, v20, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v24i32_to_v48i16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v54, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v53, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v52, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v51, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v50, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v49, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v48, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v39, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v37, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v36, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v35, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v34, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v33, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v32, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v31, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v30, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v29, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v28, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v27, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v26, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v25, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s4
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v24i32_to_v48i16:
|
|
; 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, v24
|
|
; 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 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_v24i32_to_v48i16:
|
|
; 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, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v53, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v52, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v51, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v50, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v49, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v48, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v39, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v37, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v35, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v33, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v32, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v31, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v30, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v29, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v28, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v27, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v26, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v25, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v24, v23, 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 <24 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <24 x i32> %a1 to <48 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x i32> %a to <48 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x i16> %phi
|
|
}
|
|
|
|
define inreg <48 x i16> @bitcast_v24i32_to_v48i16_scalar(<24 x i32> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v24i32_to_v48i16_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 v20, 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 v20, s30, 0
|
|
; SI-NEXT: v_mov_b32_e32 v12, s16
|
|
; SI-NEXT: v_mov_b32_e32 v13, s17
|
|
; SI-NEXT: v_mov_b32_e32 v14, s18
|
|
; SI-NEXT: v_mov_b32_e32 v15, s19
|
|
; SI-NEXT: v_mov_b32_e32 v16, s20
|
|
; SI-NEXT: v_mov_b32_e32 v17, s21
|
|
; SI-NEXT: v_writelane_b32 v20, s31, 1
|
|
; SI-NEXT: v_mov_b32_e32 v18, s22
|
|
; SI-NEXT: v_mov_b32_e32 v19, s23
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v12
|
|
; SI-NEXT: v_mov_b32_e32 v12, s24
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v13
|
|
; SI-NEXT: v_mov_b32_e32 v13, s25
|
|
; SI-NEXT: v_readfirstlane_b32 s24, v14
|
|
; SI-NEXT: v_mov_b32_e32 v14, s26
|
|
; SI-NEXT: v_readfirstlane_b32 s25, v15
|
|
; SI-NEXT: v_mov_b32_e32 v15, s27
|
|
; SI-NEXT: v_readfirstlane_b32 s22, v16
|
|
; SI-NEXT: v_mov_b32_e32 v16, s28
|
|
; SI-NEXT: v_readfirstlane_b32 s23, v17
|
|
; SI-NEXT: v_mov_b32_e32 v17, s29
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
|
|
; SI-NEXT: v_writelane_b32 v20, s34, 2
|
|
; SI-NEXT: v_readfirstlane_b32 s20, v18
|
|
; SI-NEXT: v_readfirstlane_b32 s21, v19
|
|
; SI-NEXT: v_readfirstlane_b32 s18, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s19, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s16, v14
|
|
; SI-NEXT: v_readfirstlane_b32 s17, v15
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v16
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v17
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v9
|
|
; SI-NEXT: s_and_b64 s[26:27], vcc, exec
|
|
; SI-NEXT: v_readfirstlane_b32 s5, v10
|
|
; SI-NEXT: v_writelane_b32 v20, s35, 3
|
|
; SI-NEXT: s_cbranch_scc0 .LBB13_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s88, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s90, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s91, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s92, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s94, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s95, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s41, 16
|
|
; SI-NEXT: s_lshr_b64 s[26:27], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[28:29], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[16:17], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[40:41], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB13_3
|
|
; SI-NEXT: .LBB13_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s41, s41, 3
|
|
; SI-NEXT: s_add_i32 s40, s40, 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: 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[26:27], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[28:29], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[16:17], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b32 s88, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s90, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s91, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s92, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s94, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s95, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s41, 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[40:41], 16
|
|
; SI-NEXT: .LBB13_3: ; %end
|
|
; SI-NEXT: s_lshl_b32 s27, s76, 16
|
|
; SI-NEXT: s_and_b32 s29, s40, 0xffff
|
|
; SI-NEXT: s_or_b32 s27, s29, s27
|
|
; SI-NEXT: v_mov_b32_e32 v1, s27
|
|
; SI-NEXT: s_and_b32 s27, s41, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s29, s35, 16
|
|
; SI-NEXT: s_or_b32 s27, s27, s29
|
|
; SI-NEXT: v_mov_b32_e32 v2, s27
|
|
; SI-NEXT: s_lshl_b32 s27, s74, 16
|
|
; SI-NEXT: s_and_b32 s24, s24, 0xffff
|
|
; SI-NEXT: s_or_b32 s24, s24, s27
|
|
; SI-NEXT: v_mov_b32_e32 v3, s24
|
|
; SI-NEXT: s_and_b32 s24, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s25, s34, 16
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
|
|
; SI-NEXT: s_or_b32 s24, s24, s25
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 8, v0
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s24
|
|
; SI-NEXT: s_and_b32 s22, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s24, s72, 16
|
|
; SI-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 12, v0
|
|
; SI-NEXT: s_or_b32 s22, s22, s24
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s22
|
|
; SI-NEXT: s_and_b32 s22, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s23, s31, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 16, v0
|
|
; SI-NEXT: s_or_b32 s22, s22, s23
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s22
|
|
; SI-NEXT: s_and_b32 s20, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s22, s62, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 20, v0
|
|
; SI-NEXT: s_or_b32 s20, s20, s22
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s20
|
|
; SI-NEXT: s_and_b32 s20, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s21, s30, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 24, v0
|
|
; SI-NEXT: s_or_b32 s20, s20, s21
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s20
|
|
; SI-NEXT: s_and_b32 s18, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s20, s60, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 28, v0
|
|
; SI-NEXT: s_or_b32 s18, s18, s20
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: s_and_b32 s18, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s19, s95, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 32, v0
|
|
; SI-NEXT: s_or_b32 s18, s18, s19
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: s_and_b32 s16, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s18, s58, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 36, v0
|
|
; SI-NEXT: s_or_b32 s16, s16, s18
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s16
|
|
; SI-NEXT: s_and_b32 s16, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s17, s94, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 40, v0
|
|
; SI-NEXT: s_or_b32 s16, s16, s17
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s16
|
|
; SI-NEXT: s_and_b32 s14, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s16, s56, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 44, v0
|
|
; SI-NEXT: s_or_b32 s14, s14, s16
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s14
|
|
; SI-NEXT: s_and_b32 s14, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s15, s93, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 48, v0
|
|
; SI-NEXT: s_or_b32 s14, s14, s15
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s14
|
|
; SI-NEXT: s_and_b32 s12, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s14, s46, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 52, v0
|
|
; SI-NEXT: s_or_b32 s12, s12, s14
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s12
|
|
; SI-NEXT: s_and_b32 s12, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s13, s92, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 56, v0
|
|
; SI-NEXT: s_or_b32 s12, s12, s13
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s12
|
|
; SI-NEXT: s_and_b32 s10, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s12, s44, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 60, v0
|
|
; SI-NEXT: s_or_b32 s10, s10, s12
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s10
|
|
; SI-NEXT: s_and_b32 s10, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s11, s91, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 64, v0
|
|
; SI-NEXT: s_or_b32 s10, s10, s11
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s10
|
|
; SI-NEXT: s_and_b32 s8, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s10, s42, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x44, v0
|
|
; SI-NEXT: s_or_b32 s8, s8, s10
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s8
|
|
; SI-NEXT: s_and_b32 s8, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s90, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x48, v0
|
|
; SI-NEXT: s_or_b32 s8, s8, s9
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s8
|
|
; SI-NEXT: s_and_b32 s6, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s28, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x4c, v0
|
|
; SI-NEXT: s_or_b32 s6, s6, s8
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: s_and_b32 s6, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s89, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x50, v0
|
|
; SI-NEXT: s_or_b32 s6, s6, s7
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: s_and_b32 s4, s4, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s26, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x54, v0
|
|
; SI-NEXT: s_or_b32 s4, s4, s6
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s4
|
|
; SI-NEXT: s_and_b32 s4, s5, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s88, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x58, v0
|
|
; SI-NEXT: s_or_b32 s4, s4, s5
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_mov_b32_e32 v1, s4
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: v_readlane_b32 s35, v20, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v20, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v20, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v20, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB13_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr35
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr72
|
|
; SI-NEXT: ; implicit-def: $sgpr31
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr95
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr94
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr93
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr92
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr91
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: ; implicit-def: $sgpr28
|
|
; SI-NEXT: ; implicit-def: $sgpr89
|
|
; SI-NEXT: ; implicit-def: $sgpr26
|
|
; SI-NEXT: ; implicit-def: $sgpr88
|
|
; SI-NEXT: s_branch .LBB13_2
|
|
;
|
|
; VI-LABEL: bitcast_v24i32_to_v48i16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_mov_b32_e32 v11, s16
|
|
; VI-NEXT: v_mov_b32_e32 v12, s17
|
|
; VI-NEXT: v_mov_b32_e32 v13, s18
|
|
; VI-NEXT: v_mov_b32_e32 v14, s19
|
|
; VI-NEXT: v_mov_b32_e32 v15, s20
|
|
; VI-NEXT: v_mov_b32_e32 v16, s21
|
|
; VI-NEXT: v_mov_b32_e32 v17, s22
|
|
; VI-NEXT: v_mov_b32_e32 v18, s23
|
|
; VI-NEXT: v_mov_b32_e32 v19, s24
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v11
|
|
; VI-NEXT: v_mov_b32_e32 v11, s25
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v12
|
|
; VI-NEXT: v_mov_b32_e32 v12, s26
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v13
|
|
; VI-NEXT: v_mov_b32_e32 v13, s27
|
|
; VI-NEXT: v_readfirstlane_b32 s26, v14
|
|
; VI-NEXT: v_mov_b32_e32 v14, s28
|
|
; VI-NEXT: v_readfirstlane_b32 s25, v15
|
|
; VI-NEXT: v_mov_b32_e32 v15, s29
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s24, v16
|
|
; VI-NEXT: v_readfirstlane_b32 s23, v17
|
|
; VI-NEXT: v_readfirstlane_b32 s22, v18
|
|
; VI-NEXT: v_readfirstlane_b32 s21, v19
|
|
; VI-NEXT: v_readfirstlane_b32 s20, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s19, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s18, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s17, v14
|
|
; VI-NEXT: v_readfirstlane_b32 s16, v15
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v0
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v8
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v9
|
|
; VI-NEXT: s_cbranch_scc0 .LBB13_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s27, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s28, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s29, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s43, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s44, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s42, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB13_3
|
|
; VI-NEXT: .LBB13_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s7, s7, 3
|
|
; VI-NEXT: s_add_i32 s6, s6, 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 s16, s16, 3
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s26, s26, 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_lshr_b32 s27, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s28, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s29, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s43, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s44, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s42, 16
|
|
; VI-NEXT: .LBB13_3: ; %end
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s42
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s41
|
|
; VI-NEXT: s_lshl_b32 s41, s78, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s41
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s40
|
|
; VI-NEXT: s_lshl_b32 s41, s77, 16
|
|
; VI-NEXT: s_or_b32 s40, s40, s41
|
|
; VI-NEXT: s_and_b32 s26, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s41, s76, 16
|
|
; VI-NEXT: s_or_b32 s26, s26, s41
|
|
; VI-NEXT: s_and_b32 s25, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s41, s75, 16
|
|
; VI-NEXT: s_or_b32 s25, s25, s41
|
|
; VI-NEXT: s_and_b32 s24, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s41, s74, 16
|
|
; VI-NEXT: s_or_b32 s24, s24, s41
|
|
; VI-NEXT: s_and_b32 s23, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s41, s73, 16
|
|
; VI-NEXT: s_or_b32 s23, s23, s41
|
|
; VI-NEXT: s_and_b32 s22, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s41, s72, 16
|
|
; VI-NEXT: s_or_b32 s22, s22, s41
|
|
; VI-NEXT: s_and_b32 s21, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s41, s63, 16
|
|
; VI-NEXT: s_or_b32 s21, s21, s41
|
|
; VI-NEXT: s_and_b32 s20, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s41, s62, 16
|
|
; VI-NEXT: s_or_b32 s20, s20, s41
|
|
; VI-NEXT: s_and_b32 s19, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s41, s61, 16
|
|
; VI-NEXT: s_or_b32 s19, s19, s41
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s41, s60, 16
|
|
; VI-NEXT: s_or_b32 s18, s18, s41
|
|
; VI-NEXT: s_and_b32 s17, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s59, 16
|
|
; VI-NEXT: s_or_b32 s17, s17, s41
|
|
; VI-NEXT: s_and_b32 s16, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s41, s58, 16
|
|
; VI-NEXT: s_or_b32 s16, s16, s41
|
|
; VI-NEXT: s_and_b32 s15, 0xffff, s15
|
|
; VI-NEXT: s_lshl_b32 s41, s57, 16
|
|
; VI-NEXT: s_or_b32 s15, s15, s41
|
|
; VI-NEXT: s_and_b32 s14, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s41, s56, 16
|
|
; VI-NEXT: s_or_b32 s14, s14, s41
|
|
; VI-NEXT: s_and_b32 s13, 0xffff, s13
|
|
; VI-NEXT: s_lshl_b32 s41, s47, 16
|
|
; VI-NEXT: s_or_b32 s13, s13, s41
|
|
; VI-NEXT: s_and_b32 s12, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s41, s46, 16
|
|
; VI-NEXT: s_or_b32 s12, s12, s41
|
|
; VI-NEXT: s_and_b32 s11, 0xffff, s11
|
|
; VI-NEXT: s_lshl_b32 s41, s45, 16
|
|
; VI-NEXT: s_or_b32 s11, s11, s41
|
|
; VI-NEXT: s_and_b32 s10, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s41, s44, 16
|
|
; VI-NEXT: s_or_b32 s10, s10, s41
|
|
; VI-NEXT: s_and_b32 s9, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s41, s43, 16
|
|
; VI-NEXT: s_and_b32 s8, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s29, s29, 16
|
|
; VI-NEXT: s_and_b32 s6, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s28, s28, 16
|
|
; VI-NEXT: s_and_b32 s7, 0xffff, s7
|
|
; VI-NEXT: s_lshl_b32 s27, s27, 16
|
|
; VI-NEXT: s_or_b32 s9, s9, s41
|
|
; VI-NEXT: s_or_b32 s8, s8, s29
|
|
; VI-NEXT: s_or_b32 s6, s6, s28
|
|
; VI-NEXT: s_or_b32 s7, s7, s27
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s40
|
|
; VI-NEXT: v_mov_b32_e32 v3, s26
|
|
; VI-NEXT: v_mov_b32_e32 v4, s25
|
|
; VI-NEXT: v_mov_b32_e32 v5, s24
|
|
; VI-NEXT: v_mov_b32_e32 v6, s23
|
|
; VI-NEXT: v_mov_b32_e32 v7, s22
|
|
; VI-NEXT: v_mov_b32_e32 v8, s21
|
|
; VI-NEXT: v_mov_b32_e32 v9, s20
|
|
; VI-NEXT: v_mov_b32_e32 v10, s19
|
|
; VI-NEXT: v_mov_b32_e32 v11, s18
|
|
; VI-NEXT: v_mov_b32_e32 v12, s17
|
|
; VI-NEXT: v_mov_b32_e32 v13, s16
|
|
; VI-NEXT: v_mov_b32_e32 v14, s15
|
|
; VI-NEXT: v_mov_b32_e32 v15, s14
|
|
; VI-NEXT: v_mov_b32_e32 v16, s13
|
|
; VI-NEXT: v_mov_b32_e32 v17, s12
|
|
; VI-NEXT: v_mov_b32_e32 v18, s11
|
|
; VI-NEXT: v_mov_b32_e32 v19, s10
|
|
; VI-NEXT: v_mov_b32_e32 v20, s9
|
|
; VI-NEXT: v_mov_b32_e32 v21, s8
|
|
; VI-NEXT: v_mov_b32_e32 v22, s6
|
|
; VI-NEXT: v_mov_b32_e32 v23, s7
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB13_4:
|
|
; 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: ; implicit-def: $sgpr43
|
|
; VI-NEXT: ; implicit-def: $sgpr29
|
|
; VI-NEXT: ; implicit-def: $sgpr28
|
|
; VI-NEXT: ; implicit-def: $sgpr27
|
|
; VI-NEXT: s_branch .LBB13_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v24i32_to_v48i16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s24
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v11
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s25
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s26
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v13
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s27
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v14
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s28
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v15
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s29
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v17
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v18
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v19
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s17, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s18, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s19, v15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s20, v0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s21, v1
|
|
; GFX9-NEXT: v_readfirstlane_b32 s22, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s23, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s24, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s25, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s26, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s27, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s28, v8
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_readfirstlane_b32 s29, v9
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB13_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s44, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s6, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB13_3
|
|
; GFX9-NEXT: .LBB13_2: ; %cmp.true
|
|
; 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_add_i32 s15, s15, 3
|
|
; GFX9-NEXT: s_add_i32 s14, s14, 3
|
|
; GFX9-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX9-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX9-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX9-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX9-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX9-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX9-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX9-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s44, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s6, 16
|
|
; GFX9-NEXT: .LBB13_3: ; %end
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s6, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s5, s7, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s8, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s9, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s10, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s11, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s12, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s13, s72
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s14, s63
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s15, s62
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s16, s61
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s17, s60
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s59
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s58
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s57
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s56
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s47
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s46
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s45
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s44
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s4
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s5
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s27
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB13_4:
|
|
; 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: ; implicit-def: $sgpr43
|
|
; GFX9-NEXT: ; implicit-def: $sgpr42
|
|
; GFX9-NEXT: ; implicit-def: $sgpr41
|
|
; GFX9-NEXT: ; implicit-def: $sgpr40
|
|
; GFX9-NEXT: s_branch .LBB13_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v24i32_to_v48i16_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v7, s0 :: v_dual_mov_b32 v8, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v11, s16 :: v_dual_mov_b32 v12, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v13, s18 :: v_dual_mov_b32 v14, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v15, s20 :: v_dual_mov_b32 v16, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v17, s22 :: v_dual_mov_b32 v18, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v19, s24 :: v_dual_mov_b32 v20, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v21, s26 :: v_dual_mov_b32 v22, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v23, s28 :: v_dual_mov_b32 v24, s29
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s0, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s1, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s2, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s3, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v11
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v12
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v14
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v15
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v16
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v17
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v18
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v19
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v20
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v21
|
|
; GFX11-NEXT: v_readfirstlane_b32 s15, v22
|
|
; GFX11-NEXT: v_readfirstlane_b32 s16, v23
|
|
; GFX11-NEXT: v_readfirstlane_b32 s17, v24
|
|
; GFX11-NEXT: v_readfirstlane_b32 s18, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s19, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s20, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s21, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s23, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s22, v5
|
|
; GFX11-NEXT: s_mov_b32 s74, 0
|
|
; GFX11-NEXT: s_and_b32 s24, vcc_lo, exec_lo
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB13_4
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-NEXT: s_lshr_b32 s24, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s25, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s26, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s27, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s28, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s29, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s15, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s14, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s0, 16
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s74
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB13_3
|
|
; GFX11-NEXT: .LBB13_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX11-NEXT: s_add_i32 s23, s23, 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 s15, s15, 3
|
|
; GFX11-NEXT: s_add_i32 s14, s14, 3
|
|
; GFX11-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX11-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX11-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX11-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX11-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX11-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX11-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX11-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX11-NEXT: s_add_i32 s5, s5, 3
|
|
; GFX11-NEXT: s_add_i32 s4, s4, 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 s24, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s25, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s26, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s27, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s28, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s29, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s15, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s14, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, 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, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s62
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s60
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s59
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s58
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s57
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s56
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s44
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s40
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s29
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s28
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s27
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s26
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s25
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s24
|
|
; 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, s23 :: v_dual_mov_b32 v23, s22
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB13_4:
|
|
; 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: $sgpr29
|
|
; GFX11-NEXT: ; implicit-def: $sgpr28
|
|
; GFX11-NEXT: ; implicit-def: $sgpr27
|
|
; GFX11-NEXT: ; implicit-def: $sgpr26
|
|
; GFX11-NEXT: ; implicit-def: $sgpr25
|
|
; GFX11-NEXT: ; implicit-def: $sgpr24
|
|
; 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 <24 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <24 x i32> %a1 to <48 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x i32> %a to <48 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x i16> %phi
|
|
}
|
|
|
|
define <24 x i32> @bitcast_v48i16_to_v24i32(<48 x i16> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v48i16_to_v24i32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v48, v14
|
|
; SI-NEXT: v_mov_b32_e32 v49, v12
|
|
; SI-NEXT: v_mov_b32_e32 v50, v10
|
|
; SI-NEXT: v_mov_b32_e32 v51, v8
|
|
; SI-NEXT: v_mov_b32_e32 v52, v6
|
|
; SI-NEXT: v_mov_b32_e32 v53, v4
|
|
; SI-NEXT: v_mov_b32_e32 v54, v2
|
|
; SI-NEXT: v_mov_b32_e32 v55, v0
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:68
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:60
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshlrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v43, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v42, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v41, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v29
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v8
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v12
|
|
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:52
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:44
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:40
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:36
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:196 ; 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 v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v55
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v54
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v53
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v52
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v51
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v50
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v49
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v48
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v47
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v39
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v46
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v38
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v45
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v37
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v44
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v36
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; 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: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v43
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v35
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v42
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v34
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v41
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v33
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v40
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v32
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v63
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v62
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v61
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v60
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v59
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v58
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v57
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v56
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; 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 v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v55
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v54
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v53
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v52
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v51
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v50
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v49
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v48
|
|
; 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_or_b32_e32 v0, v47, v0
|
|
; SI-NEXT: s_mov_b32 s6, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v1, v39, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v46, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v38, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v45, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v37, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v44, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v36, v7
|
|
; 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: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; 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_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v8, v43, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v35, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v42, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v34, v11
|
|
; SI-NEXT: v_or_b32_e32 v12, v41, v12
|
|
; SI-NEXT: v_or_b32_e32 v13, v33, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v40, v14
|
|
; SI-NEXT: v_or_b32_e32 v15, v32, v15
|
|
; SI-NEXT: v_or_b32_e32 v16, v63, v16
|
|
; SI-NEXT: v_or_b32_e32 v17, v62, v17
|
|
; SI-NEXT: v_or_b32_e32 v18, v61, v18
|
|
; SI-NEXT: v_or_b32_e32 v19, v60, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v59, v20
|
|
; SI-NEXT: v_or_b32_e32 v21, v58, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v57, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v56, v23
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9
|
|
; 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: v_add_i32_e32 v20, vcc, s6, v20
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v22
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 0x30000, v23
|
|
; 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 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v48i16_to_v24i32:
|
|
; 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: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_mov_b32_e32 v32, v23
|
|
; VI-NEXT: v_mov_b32_e32 v33, v22
|
|
; VI-NEXT: v_mov_b32_e32 v34, v21
|
|
; VI-NEXT: v_mov_b32_e32 v35, v20
|
|
; VI-NEXT: v_mov_b32_e32 v36, v19
|
|
; VI-NEXT: v_mov_b32_e32 v37, v18
|
|
; VI-NEXT: v_mov_b32_e32 v38, v17
|
|
; VI-NEXT: v_mov_b32_e32 v39, v16
|
|
; VI-NEXT: v_mov_b32_e32 v48, v15
|
|
; VI-NEXT: v_mov_b32_e32 v49, v14
|
|
; VI-NEXT: v_mov_b32_e32 v50, v13
|
|
; VI-NEXT: v_mov_b32_e32 v51, v12
|
|
; VI-NEXT: v_mov_b32_e32 v52, v11
|
|
; VI-NEXT: v_mov_b32_e32 v53, v10
|
|
; VI-NEXT: v_mov_b32_e32 v54, v9
|
|
; VI-NEXT: v_mov_b32_e32 v55, v8
|
|
; VI-NEXT: v_mov_b32_e32 v40, v7
|
|
; VI-NEXT: v_mov_b32_e32 v41, v6
|
|
; VI-NEXT: v_mov_b32_e32 v42, v5
|
|
; VI-NEXT: v_mov_b32_e32 v43, v4
|
|
; VI-NEXT: v_mov_b32_e32 v44, v3
|
|
; VI-NEXT: v_mov_b32_e32 v45, v2
|
|
; VI-NEXT: v_mov_b32_e32 v46, v1
|
|
; VI-NEXT: v_mov_b32_e32 v47, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 v23, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v23, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v23, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v23, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v23, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v23, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v23, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v23, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v23, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v23, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v23, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v23, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v23, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v23, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v23, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v23, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v23, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v23, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v23, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v23, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v23, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v23, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v23, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v23, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v44, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v43, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v42, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v41, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v40, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v55, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v53, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v52, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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 v23, 3
|
|
; VI-NEXT: v_add_u16_e32 v0, 3, v47
|
|
; VI-NEXT: v_add_u16_sdwa v1, v47, v23 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_add_u16_e32 v1, 3, v46
|
|
; VI-NEXT: v_add_u16_sdwa v2, v46, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v45
|
|
; VI-NEXT: v_add_u16_sdwa v3, v45, v23 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, v44
|
|
; VI-NEXT: v_add_u16_sdwa v4, v44, v23 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, v43
|
|
; VI-NEXT: v_add_u16_sdwa v5, v43, v23 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, v42
|
|
; VI-NEXT: v_add_u16_sdwa v6, v42, v23 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, v41
|
|
; VI-NEXT: v_add_u16_sdwa v7, v41, v23 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, v40
|
|
; VI-NEXT: v_add_u16_sdwa v8, v40, v23 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, v55
|
|
; VI-NEXT: v_add_u16_sdwa v9, v55, v23 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, v54
|
|
; VI-NEXT: v_add_u16_sdwa v10, v54, v23 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, v53
|
|
; VI-NEXT: v_add_u16_sdwa v11, v53, v23 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, v52
|
|
; VI-NEXT: v_add_u16_sdwa v12, v52, v23 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, v51
|
|
; VI-NEXT: v_add_u16_sdwa v13, v51, v23 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, v50
|
|
; VI-NEXT: v_add_u16_sdwa v14, v50, v23 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, v49
|
|
; VI-NEXT: v_add_u16_sdwa v15, v49, v23 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, v48
|
|
; VI-NEXT: v_add_u16_sdwa v16, v48, v23 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, v39
|
|
; VI-NEXT: v_add_u16_sdwa v17, v39, v23 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, v38
|
|
; VI-NEXT: v_add_u16_sdwa v18, v38, v23 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, v37
|
|
; VI-NEXT: v_add_u16_sdwa v19, v37, v23 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, v36
|
|
; VI-NEXT: v_add_u16_sdwa v20, v36, v23 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, v35
|
|
; VI-NEXT: v_add_u16_sdwa v21, v35, v23 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, v34
|
|
; VI-NEXT: v_add_u16_sdwa v22, v34, v23 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, v33
|
|
; VI-NEXT: v_add_u16_sdwa v24, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v22, v22, v24
|
|
; VI-NEXT: v_add_u16_e32 v24, 3, v32
|
|
; VI-NEXT: v_add_u16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB14_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48i16_to_v24i32:
|
|
; 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 v33, v22
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v13
|
|
; 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, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v12
|
|
; 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, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v11
|
|
; 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, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v10
|
|
; 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, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v1
|
|
; 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, v53
|
|
; 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, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v47
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 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:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; 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: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; 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 v9, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, 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(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, 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: .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_v48i16_to_v24i32:
|
|
; 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, v24
|
|
; 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: .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_v48i16_to_v24i32:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v24
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v53, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v55, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v64, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v52, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v51, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v50, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v49, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v48, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v39, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v38, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v37, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v36, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v35, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v34, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v33, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v32, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v31, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v30, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v29, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v28, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v27, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v26, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v25, v23, 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: .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 <48 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <48 x i16> %a1 to <24 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x i16> %a to <24 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x i32> %phi
|
|
}
|
|
|
|
define inreg <24 x i32> @bitcast_v48i16_to_v24i32_scalar(<48 x i16> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v48i16_to_v24i32_scalar:
|
|
; 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:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v61, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v63, v2
|
|
; SI-NEXT: v_mov_b32_e32 v56, v0
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: v_mov_b32_e32 v31, v22
|
|
; SI-NEXT: v_mov_b32_e32 v34, v20
|
|
; SI-NEXT: v_mov_b32_e32 v35, v18
|
|
; SI-NEXT: v_mov_b32_e32 v36, v16
|
|
; SI-NEXT: v_mov_b32_e32 v37, v14
|
|
; SI-NEXT: v_mov_b32_e32 v38, v12
|
|
; SI-NEXT: v_mov_b32_e32 v39, v10
|
|
; SI-NEXT: v_mov_b32_e32 v48, v8
|
|
; SI-NEXT: v_mov_b32_e32 v50, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v55, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v41, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v52, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v2
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v4
|
|
; SI-NEXT: s_cbranch_scc0 .LBB15_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v56
|
|
; SI-NEXT: v_or_b32_e32 v7, v0, v54
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v61
|
|
; SI-NEXT: v_or_b32_e32 v9, v0, v32
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v50
|
|
; SI-NEXT: v_or_b32_e32 v10, v0, v55
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; SI-NEXT: v_or_b32_e32 v11, v0, v62
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; SI-NEXT: v_or_b32_e32 v12, v0, v46
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; SI-NEXT: v_or_b32_e32 v13, v0, v45
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; SI-NEXT: v_or_b32_e32 v14, v0, v60
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; SI-NEXT: v_or_b32_e32 v15, v0, v59
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s17, 16
|
|
; SI-NEXT: v_or_b32_e32 v16, v0, v58
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; SI-NEXT: s_or_b32 s4, s4, s5
|
|
; SI-NEXT: s_and_b32 s5, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s19, 16
|
|
; SI-NEXT: v_or_b32_e32 v17, v0, v41
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v31
|
|
; SI-NEXT: s_or_b32 s5, s5, s6
|
|
; SI-NEXT: s_and_b32 s6, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s21, 16
|
|
; SI-NEXT: v_or_b32_e32 v18, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v24
|
|
; SI-NEXT: s_or_b32 s6, s6, s7
|
|
; SI-NEXT: s_and_b32 s7, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s23, 16
|
|
; SI-NEXT: v_or_b32_e32 v19, v0, v52
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v26
|
|
; SI-NEXT: s_or_b32 s7, s7, s8
|
|
; SI-NEXT: s_and_b32 s8, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s25, 16
|
|
; SI-NEXT: v_or_b32_e32 v20, v0, v51
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v28
|
|
; SI-NEXT: s_or_b32 s8, s8, s9
|
|
; SI-NEXT: s_and_b32 s9, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s10, s27, 16
|
|
; SI-NEXT: v_or_b32_e32 v21, v0, v29
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v30
|
|
; SI-NEXT: s_or_b32 s9, s9, s10
|
|
; SI-NEXT: s_and_b32 s10, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s11, s29, 16
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v63
|
|
; SI-NEXT: v_or_b32_e32 v22, v0, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; SI-NEXT: s_or_b32 s10, s10, s11
|
|
; SI-NEXT: v_or_b32_e32 v8, v1, v57
|
|
; SI-NEXT: v_or_b32_e32 v23, v0, v25
|
|
; SI-NEXT: v_mov_b32_e32 v0, s4
|
|
; SI-NEXT: v_mov_b32_e32 v1, s5
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: v_mov_b32_e32 v3, s7
|
|
; SI-NEXT: v_mov_b32_e32 v4, s8
|
|
; SI-NEXT: v_mov_b32_e32 v5, s9
|
|
; SI-NEXT: v_mov_b32_e32 v6, s10
|
|
; SI-NEXT: s_cbranch_execnz .LBB15_3
|
|
; SI-NEXT: .LBB15_2: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v56
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v54, v0
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v61
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v32, v0
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v50
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_or_b32_e32 v0, v55, v0
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v48
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v62, v0
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v39
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v0, v46, v0
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v38
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v45, v0
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v37
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v60, v0
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v36
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v59, v0
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v35
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v58, v0
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v34
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v41, v0
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v31
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v40, v0
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v24
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v52, v0
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v26
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v51, v0
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s17, 16
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v28
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_and_b32 s5, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s19, 16
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_or_b32 s5, s6, s5
|
|
; SI-NEXT: s_and_b32 s6, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s21, 16
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v29, v0
|
|
; SI-NEXT: s_or_b32 s6, s7, s6
|
|
; SI-NEXT: s_and_b32 s7, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s23, 16
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v30
|
|
; SI-NEXT: s_or_b32 s7, s8, s7
|
|
; SI-NEXT: s_and_b32 s8, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s25, 16
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_or_b32 s8, s9, s8
|
|
; SI-NEXT: s_and_b32 s9, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s10, s27, 16
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v27, v0
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v63
|
|
; SI-NEXT: s_or_b32 s9, s10, s9
|
|
; SI-NEXT: s_and_b32 s10, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s11, s29, 16
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v49
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: s_or_b32 s10, s11, s10
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v57, v1
|
|
; SI-NEXT: s_add_i32 s4, s4, 0x30000
|
|
; SI-NEXT: s_add_i32 s5, s5, 0x30000
|
|
; SI-NEXT: s_add_i32 s6, s6, 0x30000
|
|
; SI-NEXT: s_add_i32 s7, s7, 0x30000
|
|
; SI-NEXT: s_add_i32 s8, s8, 0x30000
|
|
; SI-NEXT: s_add_i32 s9, s9, 0x30000
|
|
; SI-NEXT: s_add_i32 s10, s10, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v0, v25, v0
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 0x30000, v1
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 0x30000, v0
|
|
; SI-NEXT: v_mov_b32_e32 v0, s4
|
|
; SI-NEXT: v_mov_b32_e32 v1, s5
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: v_mov_b32_e32 v3, s7
|
|
; SI-NEXT: v_mov_b32_e32 v4, s8
|
|
; SI-NEXT: v_mov_b32_e32 v5, s9
|
|
; SI-NEXT: v_mov_b32_e32 v6, s10
|
|
; SI-NEXT: .LBB15_3: ; %end
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB15_4:
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v46, v51
|
|
; SI-NEXT: v_mov_b32_e32 v51, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v34
|
|
; SI-NEXT: v_mov_b32_e32 v34, v30
|
|
; SI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_mov_b32_e32 v47, v52
|
|
; SI-NEXT: v_mov_b32_e32 v52, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v35
|
|
; SI-NEXT: v_mov_b32_e32 v35, v28
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v55, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v41
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v41, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v26
|
|
; SI-NEXT: v_mov_b32_e32 v42, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v24
|
|
; SI-NEXT: v_mov_b32_e32 v33, v32
|
|
; SI-NEXT: v_mov_b32_e32 v32, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v40
|
|
; SI-NEXT: v_mov_b32_e32 v40, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v31
|
|
; SI-NEXT: v_mov_b32_e32 v43, v25
|
|
; SI-NEXT: v_mov_b32_e32 v44, v27
|
|
; SI-NEXT: v_mov_b32_e32 v53, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v29
|
|
; 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: v_mov_b32_e32 v26, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v55
|
|
; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v30, v34
|
|
; SI-NEXT: v_mov_b32_e32 v34, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v51
|
|
; SI-NEXT: v_mov_b32_e32 v51, v46
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v29, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v53
|
|
; SI-NEXT: v_mov_b32_e32 v27, v44
|
|
; SI-NEXT: v_mov_b32_e32 v25, v43
|
|
; SI-NEXT: v_mov_b32_e32 v31, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v40
|
|
; SI-NEXT: v_mov_b32_e32 v40, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v32
|
|
; SI-NEXT: v_mov_b32_e32 v32, v33
|
|
; SI-NEXT: v_mov_b32_e32 v24, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v42
|
|
; SI-NEXT: v_mov_b32_e32 v28, v35
|
|
; SI-NEXT: v_mov_b32_e32 v35, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v52
|
|
; SI-NEXT: v_mov_b32_e32 v52, v47
|
|
; SI-NEXT: s_branch .LBB15_2
|
|
;
|
|
; VI-LABEL: bitcast_v48i16_to_v24i32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v32, v9
|
|
; VI-NEXT: v_mov_b32_e32 v33, v8
|
|
; VI-NEXT: v_mov_b32_e32 v34, v7
|
|
; VI-NEXT: v_mov_b32_e32 v35, v6
|
|
; VI-NEXT: v_mov_b32_e32 v36, v5
|
|
; VI-NEXT: v_mov_b32_e32 v37, v4
|
|
; VI-NEXT: v_mov_b32_e32 v38, v3
|
|
; VI-NEXT: v_mov_b32_e32 v39, v2
|
|
; VI-NEXT: v_mov_b32_e32 v48, v1
|
|
; VI-NEXT: v_mov_b32_e32 v49, v0
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; 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, s43, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s44, s42, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s44
|
|
; VI-NEXT: s_and_b32 s44, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s45, s41, 16
|
|
; VI-NEXT: s_or_b32 s44, s44, s45
|
|
; VI-NEXT: s_and_b32 s45, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s46, s40, 16
|
|
; VI-NEXT: s_or_b32 s45, s45, s46
|
|
; VI-NEXT: s_and_b32 s46, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s47, s15, 16
|
|
; VI-NEXT: s_or_b32 s46, s46, s47
|
|
; VI-NEXT: s_and_b32 s47, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s56, s14, 16
|
|
; VI-NEXT: v_mov_b32_e32 v0, 16
|
|
; VI-NEXT: s_or_b32 s47, s47, s56
|
|
; VI-NEXT: s_and_b32 s56, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s57, s13, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s56, s56, s57
|
|
; VI-NEXT: s_and_b32 s57, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s58, s12, 16
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s57, s57, s58
|
|
; VI-NEXT: s_and_b32 s58, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s59, s11, 16
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s58, s58, s59
|
|
; VI-NEXT: s_and_b32 s59, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s60, s10, 16
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s59, s59, s60
|
|
; VI-NEXT: s_and_b32 s60, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s61, s9, 16
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s60, s60, s61
|
|
; VI-NEXT: s_and_b32 s61, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s62, s8, 16
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s61, s61, s62
|
|
; VI-NEXT: s_and_b32 s62, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s63, s7, 16
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s62, s62, s63
|
|
; VI-NEXT: s_and_b32 s63, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s72, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s63, s63, s72
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v0, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s44
|
|
; VI-NEXT: v_mov_b32_e32 v3, s45
|
|
; VI-NEXT: v_mov_b32_e32 v4, s46
|
|
; VI-NEXT: v_mov_b32_e32 v5, s47
|
|
; VI-NEXT: v_mov_b32_e32 v6, s56
|
|
; VI-NEXT: v_mov_b32_e32 v7, s57
|
|
; VI-NEXT: v_mov_b32_e32 v8, s58
|
|
; VI-NEXT: v_mov_b32_e32 v9, s59
|
|
; VI-NEXT: v_mov_b32_e32 v10, s60
|
|
; VI-NEXT: v_mov_b32_e32 v11, s61
|
|
; VI-NEXT: v_mov_b32_e32 v12, s62
|
|
; VI-NEXT: v_mov_b32_e32 v13, s63
|
|
; VI-NEXT: s_cbranch_execnz .LBB15_3
|
|
; VI-NEXT: .LBB15_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v1, 16
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v49
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v48
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s43, 16
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: v_add_u32_e32 v15, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v39
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_and_b32 s5, s17, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s16, s42, 16
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s5, s16, s5
|
|
; VI-NEXT: s_and_b32 s16, s18, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s17, s41, 16
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v38
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s16, s17, s16
|
|
; VI-NEXT: s_and_b32 s17, s19, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s18, s40, 16
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s17, s18, s17
|
|
; VI-NEXT: s_and_b32 s18, s20, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s15, s15, 16
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: v_add_u32_e32 v17, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v37
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s15, s15, s18
|
|
; VI-NEXT: s_and_b32 s18, s21, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s14, s14, 16
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s14, s14, s18
|
|
; VI-NEXT: s_and_b32 s18, s22, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s13, s13, 16
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v36
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s13, s13, s18
|
|
; VI-NEXT: s_and_b32 s18, s23, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s12, s12, 16
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s12, s12, s18
|
|
; VI-NEXT: s_and_b32 s18, s24, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s11, s11, 16
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: v_add_u32_e32 v19, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v35
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s11, s11, s18
|
|
; VI-NEXT: s_and_b32 s18, s25, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s10, s10, 16
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s10, s10, s18
|
|
; VI-NEXT: s_and_b32 s18, s26, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s9, s9, 16
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v34
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s9, s9, s18
|
|
; VI-NEXT: s_and_b32 s18, s27, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s8, s8, 16
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s8, s8, s18
|
|
; VI-NEXT: s_and_b32 s18, s28, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s7, s7, 16
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: v_add_u32_e32 v21, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v33
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s7, s7, s18
|
|
; VI-NEXT: s_and_b32 s18, s29, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s6, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s6, s6, s18
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v32
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v1, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_add_i32 s4, s4, 0x30000
|
|
; VI-NEXT: s_add_i32 s5, s5, 0x30000
|
|
; VI-NEXT: s_add_i32 s16, s16, 0x30000
|
|
; VI-NEXT: s_add_i32 s17, s17, 0x30000
|
|
; VI-NEXT: s_add_i32 s15, s15, 0x30000
|
|
; VI-NEXT: s_add_i32 s14, s14, 0x30000
|
|
; VI-NEXT: s_add_i32 s13, s13, 0x30000
|
|
; VI-NEXT: s_add_i32 s12, s12, 0x30000
|
|
; VI-NEXT: s_add_i32 s11, s11, 0x30000
|
|
; VI-NEXT: s_add_i32 s10, s10, 0x30000
|
|
; VI-NEXT: s_add_i32 s9, s9, 0x30000
|
|
; VI-NEXT: s_add_i32 s8, s8, 0x30000
|
|
; VI-NEXT: s_add_i32 s7, s7, 0x30000
|
|
; VI-NEXT: s_add_i32 s6, s6, 0x30000
|
|
; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: v_add_u32_e32 v23, vcc, 0x30000, v0
|
|
; 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, s15
|
|
; VI-NEXT: v_mov_b32_e32 v5, s14
|
|
; VI-NEXT: v_mov_b32_e32 v6, s13
|
|
; VI-NEXT: v_mov_b32_e32 v7, s12
|
|
; VI-NEXT: v_mov_b32_e32 v8, s11
|
|
; VI-NEXT: v_mov_b32_e32 v9, s10
|
|
; VI-NEXT: v_mov_b32_e32 v10, s9
|
|
; VI-NEXT: v_mov_b32_e32 v11, s8
|
|
; VI-NEXT: v_mov_b32_e32 v12, s7
|
|
; VI-NEXT: v_mov_b32_e32 v13, s6
|
|
; VI-NEXT: .LBB15_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB15_4:
|
|
; 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_branch .LBB15_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v48i16_to_v24i32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v0
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, 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: buffer_store_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v32
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v33
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v49
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s17, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s18, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s19, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s20, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s21, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s22, s12
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s23, s13
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s24, s14
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s25, s15
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s29, s40
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB15_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v33
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s19
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB15_3
|
|
; GFX9-NEXT: .LBB15_2: ; %cmp.true
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v49
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v48
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v39
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v38
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v37
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v36
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v35
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v34
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v33
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v23
|
|
; GFX9-NEXT: v_pk_add_u16 v0, s6, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, s7, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, s8, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, s9, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, s10, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, s11, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, s12, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, s13, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, s14, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v9, s15, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, s16, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, s17, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, s18, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, s19, 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: .LBB15_3: ; %end
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB15_4:
|
|
; 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_branch .LBB15_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v24i32_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB15_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; 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, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB15_3
|
|
; GFX11-TRUE16-NEXT: .LBB15_2: ; %cmp.true
|
|
; 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, s4, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s17, 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: .LBB15_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB15_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB15_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v48i16_to_v24i32_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s4, s0, s4
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s44
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s16, s5
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s17, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s18, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s19, s8
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s20, s9
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s21, s10
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s22, s11
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s23, s12
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s24, s13
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s25, s14
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s26, s43
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB15_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB15_3
|
|
; GFX11-FAKE16-NEXT: .LBB15_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, s4, 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 v2, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, s17, 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: .LBB15_3: ; %end
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB15_4:
|
|
; GFX11-FAKE16-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
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB15_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <48 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <48 x i16> %a1 to <24 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x i16> %a to <24 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x i32> %phi
|
|
}
|
|
|
|
define <48 x half> @bitcast_v24i32_to_v48f16(<24 x i32> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v24i32_to_v48f16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v25
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; 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: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; 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: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; 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_lshrrev_b32_e32 v25, 16, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v1
|
|
; SI-NEXT: ; implicit-def: $vgpr1
|
|
; SI-NEXT: ; implicit-def: $vgpr2
|
|
; SI-NEXT: ; implicit-def: $vgpr3
|
|
; SI-NEXT: ; implicit-def: $vgpr4
|
|
; SI-NEXT: ; implicit-def: $vgpr5
|
|
; SI-NEXT: ; implicit-def: $vgpr6
|
|
; 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: $vgpr13
|
|
; SI-NEXT: ; implicit-def: $vgpr14
|
|
; SI-NEXT: ; implicit-def: $vgpr15
|
|
; SI-NEXT: ; implicit-def: $vgpr16
|
|
; SI-NEXT: ; implicit-def: $vgpr17
|
|
; SI-NEXT: ; implicit-def: $vgpr18
|
|
; SI-NEXT: ; implicit-def: $vgpr19
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr22
|
|
; SI-NEXT: ; implicit-def: $vgpr23
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v2, vcc, 3, v2
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v1
|
|
; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v3
|
|
; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v5
|
|
; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v7
|
|
; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v9
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v11
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v13
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v15
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v51
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v49
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v36
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v48
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v52
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v40
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v45
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v47
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v61
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v33
|
|
; SI-NEXT: v_mov_b32_e32 v35, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: .LBB16_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v30
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v3
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
|
|
; SI-NEXT: v_or_b32_e32 v2, v4, v2
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v29
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v28
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v26
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v63
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v25
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v61
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v62
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v59
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v60
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v57
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v58
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v56
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v46
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v43
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v44
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v42
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v41
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v55
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v50
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v53
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v48
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v38
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v36
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v34
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v37
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v39
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v49
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v51
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; 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_v24i32_to_v48f16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; VI-NEXT: .LBB16_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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_lshlrev_b32_e32 v27, 16, v27
|
|
; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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_or_b32_sdwa v20, v20, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v24i32_to_v48f16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v54, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v53, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v52, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v51, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v50, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v49, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v48, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v39, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v37, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v36, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v35, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v34, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v33, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v32, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v31, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v30, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v29, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v28, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v27, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v26, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v25, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s4
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v24i32_to_v48f16:
|
|
; 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, v24
|
|
; 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 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_v24i32_to_v48f16:
|
|
; 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, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v53, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v52, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v51, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v50, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v49, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v48, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v39, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v37, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v35, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v33, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v32, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v31, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v30, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v29, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v28, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v27, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v26, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v25, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v24, v23, 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 <24 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <24 x i32> %a1 to <48 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x i32> %a to <48 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x half> %phi
|
|
}
|
|
|
|
define inreg <48 x half> @bitcast_v24i32_to_v48f16_scalar(<24 x i32> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v24i32_to_v48f16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v12, s16
|
|
; SI-NEXT: v_mov_b32_e32 v13, s17
|
|
; SI-NEXT: v_mov_b32_e32 v14, s18
|
|
; SI-NEXT: v_mov_b32_e32 v15, s19
|
|
; SI-NEXT: v_mov_b32_e32 v16, s20
|
|
; SI-NEXT: v_mov_b32_e32 v17, s21
|
|
; SI-NEXT: v_mov_b32_e32 v18, s22
|
|
; SI-NEXT: v_mov_b32_e32 v19, s23
|
|
; SI-NEXT: v_readfirstlane_b32 s23, v12
|
|
; SI-NEXT: v_mov_b32_e32 v12, s24
|
|
; SI-NEXT: v_readfirstlane_b32 s24, v13
|
|
; SI-NEXT: v_mov_b32_e32 v13, s25
|
|
; SI-NEXT: v_readfirstlane_b32 s25, v14
|
|
; SI-NEXT: v_mov_b32_e32 v14, s26
|
|
; SI-NEXT: v_readfirstlane_b32 s26, v15
|
|
; SI-NEXT: v_mov_b32_e32 v15, s27
|
|
; SI-NEXT: v_readfirstlane_b32 s27, v16
|
|
; SI-NEXT: v_mov_b32_e32 v16, s28
|
|
; SI-NEXT: v_readfirstlane_b32 s28, v17
|
|
; SI-NEXT: v_mov_b32_e32 v17, s29
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s29, v18
|
|
; SI-NEXT: v_readfirstlane_b32 s22, v19
|
|
; SI-NEXT: v_readfirstlane_b32 s21, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s20, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s19, v14
|
|
; SI-NEXT: v_readfirstlane_b32 s18, v15
|
|
; SI-NEXT: v_readfirstlane_b32 s17, v16
|
|
; SI-NEXT: v_readfirstlane_b32 s16, v17
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v9
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_cbranch_scc0 .LBB17_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s4, s9, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s6, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s7, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s8, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s10, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s11, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s12, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s13, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s14, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s15, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s16, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s17, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s18, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s19, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s20, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s21, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s22, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s29, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s28, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s27, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s26, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s25, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s24, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s23, 16
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, s4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, s19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, s20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, s21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, s22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, s29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, s28
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, s27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, s26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, s25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, s24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, s23
|
|
; SI-NEXT: s_cbranch_execnz .LBB17_3
|
|
; SI-NEXT: .LBB17_2: ; %cmp.true
|
|
; SI-NEXT: s_add_i32 s23, s23, 3
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: s_add_i32 s29, s29, 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: 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 s8, s8, 3
|
|
; SI-NEXT: s_add_i32 s7, s7, 3
|
|
; SI-NEXT: s_add_i32 s6, s6, 3
|
|
; SI-NEXT: s_add_i32 s9, s9, 3
|
|
; SI-NEXT: s_lshr_b32 s4, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s5, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s40, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s41, s26, 16
|
|
; SI-NEXT: s_lshr_b32 s42, s27, 16
|
|
; SI-NEXT: s_lshr_b32 s43, s28, 16
|
|
; SI-NEXT: s_lshr_b32 s44, s29, 16
|
|
; SI-NEXT: s_lshr_b32 s45, s22, 16
|
|
; SI-NEXT: s_lshr_b32 s46, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s47, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s56, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s57, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s58, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s59, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s60, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s61, s14, 16
|
|
; SI-NEXT: s_lshr_b32 s62, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s63, s12, 16
|
|
; SI-NEXT: s_lshr_b32 s72, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s73, s10, 16
|
|
; SI-NEXT: s_lshr_b32 s74, s8, 16
|
|
; SI-NEXT: s_lshr_b32 s75, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s76, s6, 16
|
|
; SI-NEXT: s_lshr_b32 s77, s9, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, s19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, s20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, s21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, s22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, s29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, s28
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, s27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, s26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, s25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, s24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, s23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s77
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s76
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s75
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s74
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s73
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s72
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s62
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s61
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s60
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s58
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s56
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, s47
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s46
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, s45
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, s44
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, s43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, s42
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, s41
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, s40
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, s5
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, s4
|
|
; SI-NEXT: .LBB17_3: ; %end
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v40, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v55
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v53
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, v51
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v49
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v39
|
|
; SI-NEXT: v_or_b32_e32 v55, v55, v40
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v48
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v37
|
|
; SI-NEXT: buffer_store_dword v55, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v55, vcc, 4, v0
|
|
; SI-NEXT: v_or_b32_e32 v53, v53, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v38
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, v35
|
|
; SI-NEXT: buffer_store_dword v53, v55, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v53, vcc, 8, v0
|
|
; SI-NEXT: v_or_b32_e32 v51, v51, v52
|
|
; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v36
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v33
|
|
; SI-NEXT: buffer_store_dword v51, v53, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v51, vcc, 12, v0
|
|
; SI-NEXT: v_or_b32_e32 v49, v49, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, v34
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
|
|
; SI-NEXT: buffer_store_dword v49, v51, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v49, vcc, 16, v0
|
|
; SI-NEXT: v_or_b32_e32 v39, v48, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: buffer_store_dword v39, v49, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v39, vcc, 20, v0
|
|
; SI-NEXT: v_or_b32_e32 v37, v38, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v30, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: buffer_store_dword v37, v39, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v37, vcc, 24, v0
|
|
; SI-NEXT: v_or_b32_e32 v35, v36, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: buffer_store_dword v35, v37, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v35, vcc, 28, v0
|
|
; SI-NEXT: v_or_b32_e32 v33, v34, v33
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: buffer_store_dword v33, v35, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v33, vcc, 32, v0
|
|
; SI-NEXT: v_or_b32_e32 v31, v32, v31
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: buffer_store_dword v31, v33, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v31, vcc, 36, v0
|
|
; SI-NEXT: v_or_b32_e32 v29, v30, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
|
|
; SI-NEXT: buffer_store_dword v29, v31, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v29, vcc, 40, v0
|
|
; SI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; SI-NEXT: buffer_store_dword v27, v29, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 44, v0
|
|
; SI-NEXT: v_or_b32_e32 v24, v26, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v25
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v23
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 52, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v22, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v21
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 56, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v19
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 60, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v18, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v17
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 64, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v16, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v15
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 0x44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v13
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 0x48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v12, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v11
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 0x4c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v10, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: buffer_store_dword v7, v11, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v9
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 0x50, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_or_b32_e32 v5, v7, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: buffer_store_dword v5, v9, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v8
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 0x54, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v3, v5, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v6
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 0x58, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v4
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB17_4:
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr22
|
|
; SI-NEXT: ; implicit-def: $vgpr23
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr18
|
|
; SI-NEXT: ; implicit-def: $vgpr19
|
|
; SI-NEXT: ; implicit-def: $vgpr16
|
|
; SI-NEXT: ; implicit-def: $vgpr17
|
|
; SI-NEXT: ; implicit-def: $vgpr14
|
|
; SI-NEXT: ; implicit-def: $vgpr15
|
|
; SI-NEXT: ; implicit-def: $vgpr12
|
|
; SI-NEXT: ; implicit-def: $vgpr13
|
|
; SI-NEXT: ; implicit-def: $vgpr10
|
|
; SI-NEXT: ; implicit-def: $vgpr11
|
|
; SI-NEXT: ; implicit-def: $vgpr7
|
|
; SI-NEXT: ; implicit-def: $vgpr9
|
|
; SI-NEXT: ; implicit-def: $vgpr5
|
|
; SI-NEXT: ; implicit-def: $vgpr8
|
|
; SI-NEXT: ; implicit-def: $vgpr3
|
|
; SI-NEXT: ; implicit-def: $vgpr6
|
|
; SI-NEXT: ; implicit-def: $vgpr2
|
|
; SI-NEXT: ; implicit-def: $vgpr4
|
|
; SI-NEXT: ; implicit-def: $vgpr1
|
|
; SI-NEXT: s_branch .LBB17_2
|
|
;
|
|
; VI-LABEL: bitcast_v24i32_to_v48f16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_mov_b32_e32 v11, s16
|
|
; VI-NEXT: v_mov_b32_e32 v12, s17
|
|
; VI-NEXT: v_mov_b32_e32 v13, s18
|
|
; VI-NEXT: v_mov_b32_e32 v14, s19
|
|
; VI-NEXT: v_mov_b32_e32 v15, s20
|
|
; VI-NEXT: v_mov_b32_e32 v16, s21
|
|
; VI-NEXT: v_mov_b32_e32 v17, s22
|
|
; VI-NEXT: v_mov_b32_e32 v18, s23
|
|
; VI-NEXT: v_mov_b32_e32 v19, s24
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v11
|
|
; VI-NEXT: v_mov_b32_e32 v11, s25
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v12
|
|
; VI-NEXT: v_mov_b32_e32 v12, s26
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v13
|
|
; VI-NEXT: v_mov_b32_e32 v13, s27
|
|
; VI-NEXT: v_readfirstlane_b32 s26, v14
|
|
; VI-NEXT: v_mov_b32_e32 v14, s28
|
|
; VI-NEXT: v_readfirstlane_b32 s25, v15
|
|
; VI-NEXT: v_mov_b32_e32 v15, s29
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s24, v16
|
|
; VI-NEXT: v_readfirstlane_b32 s23, v17
|
|
; VI-NEXT: v_readfirstlane_b32 s22, v18
|
|
; VI-NEXT: v_readfirstlane_b32 s21, v19
|
|
; VI-NEXT: v_readfirstlane_b32 s20, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s19, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s18, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s17, v14
|
|
; VI-NEXT: v_readfirstlane_b32 s16, v15
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v0
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v8
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v9
|
|
; VI-NEXT: s_cbranch_scc0 .LBB17_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s27, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s28, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s29, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s43, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s44, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s42, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB17_3
|
|
; VI-NEXT: .LBB17_2: ; %cmp.true
|
|
; VI-NEXT: s_add_i32 s7, s7, 3
|
|
; VI-NEXT: s_add_i32 s6, s6, 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 s16, s16, 3
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s26, s26, 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_lshr_b32 s27, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s28, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s29, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s43, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s44, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s42, 16
|
|
; VI-NEXT: .LBB17_3: ; %end
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s42
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s41
|
|
; VI-NEXT: s_lshl_b32 s41, s78, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s41
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s40
|
|
; VI-NEXT: s_lshl_b32 s41, s77, 16
|
|
; VI-NEXT: s_or_b32 s40, s40, s41
|
|
; VI-NEXT: s_and_b32 s26, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s41, s76, 16
|
|
; VI-NEXT: s_or_b32 s26, s26, s41
|
|
; VI-NEXT: s_and_b32 s25, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s41, s75, 16
|
|
; VI-NEXT: s_or_b32 s25, s25, s41
|
|
; VI-NEXT: s_and_b32 s24, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s41, s74, 16
|
|
; VI-NEXT: s_or_b32 s24, s24, s41
|
|
; VI-NEXT: s_and_b32 s23, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s41, s73, 16
|
|
; VI-NEXT: s_or_b32 s23, s23, s41
|
|
; VI-NEXT: s_and_b32 s22, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s41, s72, 16
|
|
; VI-NEXT: s_or_b32 s22, s22, s41
|
|
; VI-NEXT: s_and_b32 s21, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s41, s63, 16
|
|
; VI-NEXT: s_or_b32 s21, s21, s41
|
|
; VI-NEXT: s_and_b32 s20, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s41, s62, 16
|
|
; VI-NEXT: s_or_b32 s20, s20, s41
|
|
; VI-NEXT: s_and_b32 s19, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s41, s61, 16
|
|
; VI-NEXT: s_or_b32 s19, s19, s41
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s41, s60, 16
|
|
; VI-NEXT: s_or_b32 s18, s18, s41
|
|
; VI-NEXT: s_and_b32 s17, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s59, 16
|
|
; VI-NEXT: s_or_b32 s17, s17, s41
|
|
; VI-NEXT: s_and_b32 s16, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s41, s58, 16
|
|
; VI-NEXT: s_or_b32 s16, s16, s41
|
|
; VI-NEXT: s_and_b32 s15, 0xffff, s15
|
|
; VI-NEXT: s_lshl_b32 s41, s57, 16
|
|
; VI-NEXT: s_or_b32 s15, s15, s41
|
|
; VI-NEXT: s_and_b32 s14, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s41, s56, 16
|
|
; VI-NEXT: s_or_b32 s14, s14, s41
|
|
; VI-NEXT: s_and_b32 s13, 0xffff, s13
|
|
; VI-NEXT: s_lshl_b32 s41, s47, 16
|
|
; VI-NEXT: s_or_b32 s13, s13, s41
|
|
; VI-NEXT: s_and_b32 s12, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s41, s46, 16
|
|
; VI-NEXT: s_or_b32 s12, s12, s41
|
|
; VI-NEXT: s_and_b32 s11, 0xffff, s11
|
|
; VI-NEXT: s_lshl_b32 s41, s45, 16
|
|
; VI-NEXT: s_or_b32 s11, s11, s41
|
|
; VI-NEXT: s_and_b32 s10, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s41, s44, 16
|
|
; VI-NEXT: s_or_b32 s10, s10, s41
|
|
; VI-NEXT: s_and_b32 s9, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s41, s43, 16
|
|
; VI-NEXT: s_and_b32 s8, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s29, s29, 16
|
|
; VI-NEXT: s_and_b32 s6, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s28, s28, 16
|
|
; VI-NEXT: s_and_b32 s7, 0xffff, s7
|
|
; VI-NEXT: s_lshl_b32 s27, s27, 16
|
|
; VI-NEXT: s_or_b32 s9, s9, s41
|
|
; VI-NEXT: s_or_b32 s8, s8, s29
|
|
; VI-NEXT: s_or_b32 s6, s6, s28
|
|
; VI-NEXT: s_or_b32 s7, s7, s27
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s40
|
|
; VI-NEXT: v_mov_b32_e32 v3, s26
|
|
; VI-NEXT: v_mov_b32_e32 v4, s25
|
|
; VI-NEXT: v_mov_b32_e32 v5, s24
|
|
; VI-NEXT: v_mov_b32_e32 v6, s23
|
|
; VI-NEXT: v_mov_b32_e32 v7, s22
|
|
; VI-NEXT: v_mov_b32_e32 v8, s21
|
|
; VI-NEXT: v_mov_b32_e32 v9, s20
|
|
; VI-NEXT: v_mov_b32_e32 v10, s19
|
|
; VI-NEXT: v_mov_b32_e32 v11, s18
|
|
; VI-NEXT: v_mov_b32_e32 v12, s17
|
|
; VI-NEXT: v_mov_b32_e32 v13, s16
|
|
; VI-NEXT: v_mov_b32_e32 v14, s15
|
|
; VI-NEXT: v_mov_b32_e32 v15, s14
|
|
; VI-NEXT: v_mov_b32_e32 v16, s13
|
|
; VI-NEXT: v_mov_b32_e32 v17, s12
|
|
; VI-NEXT: v_mov_b32_e32 v18, s11
|
|
; VI-NEXT: v_mov_b32_e32 v19, s10
|
|
; VI-NEXT: v_mov_b32_e32 v20, s9
|
|
; VI-NEXT: v_mov_b32_e32 v21, s8
|
|
; VI-NEXT: v_mov_b32_e32 v22, s6
|
|
; VI-NEXT: v_mov_b32_e32 v23, s7
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB17_4:
|
|
; 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: ; implicit-def: $sgpr43
|
|
; VI-NEXT: ; implicit-def: $sgpr29
|
|
; VI-NEXT: ; implicit-def: $sgpr28
|
|
; VI-NEXT: ; implicit-def: $sgpr27
|
|
; VI-NEXT: s_branch .LBB17_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v24i32_to_v48f16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s24
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v11
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s25
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s26
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v13
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s27
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v14
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s28
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v15
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s29
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v17
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v18
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v19
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s17, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s18, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s19, v15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s20, v0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s21, v1
|
|
; GFX9-NEXT: v_readfirstlane_b32 s22, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s23, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s24, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s25, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s26, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s27, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s28, v8
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_readfirstlane_b32 s29, v9
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB17_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s44, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s6, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB17_3
|
|
; GFX9-NEXT: .LBB17_2: ; %cmp.true
|
|
; 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_add_i32 s15, s15, 3
|
|
; GFX9-NEXT: s_add_i32 s14, s14, 3
|
|
; GFX9-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX9-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX9-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX9-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX9-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX9-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX9-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX9-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s44, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s6, 16
|
|
; GFX9-NEXT: .LBB17_3: ; %end
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s6, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s5, s7, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s8, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s9, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s10, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s11, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s12, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s13, s72
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s14, s63
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s15, s62
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s16, s61
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s17, s60
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s59
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s58
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s57
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s56
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s47
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s46
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s45
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s44
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s4
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s5
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s27
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB17_4:
|
|
; 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: ; implicit-def: $sgpr43
|
|
; GFX9-NEXT: ; implicit-def: $sgpr42
|
|
; GFX9-NEXT: ; implicit-def: $sgpr41
|
|
; GFX9-NEXT: ; implicit-def: $sgpr40
|
|
; GFX9-NEXT: s_branch .LBB17_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v24i32_to_v48f16_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v7, s0 :: v_dual_mov_b32 v8, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v11, s16 :: v_dual_mov_b32 v12, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v13, s18 :: v_dual_mov_b32 v14, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v15, s20 :: v_dual_mov_b32 v16, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v17, s22 :: v_dual_mov_b32 v18, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v19, s24 :: v_dual_mov_b32 v20, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v21, s26 :: v_dual_mov_b32 v22, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v23, s28 :: v_dual_mov_b32 v24, s29
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s0, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s1, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s2, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s3, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v11
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v12
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v14
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v15
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v16
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v17
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v18
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v19
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v20
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v21
|
|
; GFX11-NEXT: v_readfirstlane_b32 s15, v22
|
|
; GFX11-NEXT: v_readfirstlane_b32 s16, v23
|
|
; GFX11-NEXT: v_readfirstlane_b32 s17, v24
|
|
; GFX11-NEXT: v_readfirstlane_b32 s18, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s19, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s20, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s21, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s23, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s22, v5
|
|
; GFX11-NEXT: s_mov_b32 s74, 0
|
|
; GFX11-NEXT: s_and_b32 s24, vcc_lo, exec_lo
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB17_4
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-NEXT: s_lshr_b32 s24, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s25, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s26, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s27, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s28, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s29, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s15, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s14, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s0, 16
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s74
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB17_3
|
|
; GFX11-NEXT: .LBB17_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_i32 s22, s22, 3
|
|
; GFX11-NEXT: s_add_i32 s23, s23, 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 s15, s15, 3
|
|
; GFX11-NEXT: s_add_i32 s14, s14, 3
|
|
; GFX11-NEXT: s_add_i32 s13, s13, 3
|
|
; GFX11-NEXT: s_add_i32 s12, s12, 3
|
|
; GFX11-NEXT: s_add_i32 s11, s11, 3
|
|
; GFX11-NEXT: s_add_i32 s10, s10, 3
|
|
; GFX11-NEXT: s_add_i32 s9, s9, 3
|
|
; GFX11-NEXT: s_add_i32 s8, s8, 3
|
|
; GFX11-NEXT: s_add_i32 s7, s7, 3
|
|
; GFX11-NEXT: s_add_i32 s6, s6, 3
|
|
; GFX11-NEXT: s_add_i32 s5, s5, 3
|
|
; GFX11-NEXT: s_add_i32 s4, s4, 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 s24, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s25, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s26, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s27, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s28, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s29, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s15, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s14, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, 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, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s62
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s60
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s59
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s58
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s57
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s56
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s44
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s40
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s29
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s28
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s27
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s26
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s25
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s24
|
|
; 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, s23 :: v_dual_mov_b32 v23, s22
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB17_4:
|
|
; 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: $sgpr29
|
|
; GFX11-NEXT: ; implicit-def: $sgpr28
|
|
; GFX11-NEXT: ; implicit-def: $sgpr27
|
|
; GFX11-NEXT: ; implicit-def: $sgpr26
|
|
; GFX11-NEXT: ; implicit-def: $sgpr25
|
|
; GFX11-NEXT: ; implicit-def: $sgpr24
|
|
; 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 <24 x i32> %a, splat (i32 3)
|
|
%a2 = bitcast <24 x i32> %a1 to <48 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x i32> %a to <48 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x half> %phi
|
|
}
|
|
|
|
define <24 x i32> @bitcast_v48f16_to_v24i32(<48 x half> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v48f16_to_v24i32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v14
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:24
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:20
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:32
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:28
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:48
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:44
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:56
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:52
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:64
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:60
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v1
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, v5
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v6
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v11
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, v12
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v15
|
|
; 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_cvt_f16_f32_e32 v0, v20
|
|
; 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_cvt_f16_f32_e32 v0, v23
|
|
; 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_cvt_f16_f32_e32 v0, v22
|
|
; 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_cvt_f16_f32_e32 v0, v25
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v24
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v27
|
|
; 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_cvt_f16_f32_e32 v0, v26
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v29
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v28
|
|
; 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_cvt_f16_f32_e32 v0, v30
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v31
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v40, v63
|
|
; 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_cvt_f16_f32_e32 v0, v41
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v62
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v42
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, v61
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v43
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v60
|
|
; 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_cvt_f16_f32_e32 v0, v44
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v59
|
|
; 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_cvt_f16_f32_e32 v0, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v45, v58
|
|
; 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_cvt_f16_f32_e32 v0, v46
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, v57
|
|
; 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_cvt_f16_f32_e32 v0, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v56
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 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 v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v33
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v55
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v49
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v45
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v43
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v41
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_or_b32_e32 v0, v54, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v52, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v50, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v48, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v38, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v36, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v34, v6
|
|
; SI-NEXT: v_or_b32_e32 v21, v44, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v42, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v40, v23
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; 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: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; 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: buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v20, v46, v20
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; 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: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v55
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v53
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v52
|
|
; 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, v51
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v48
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v38
|
|
; 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, v49
|
|
; 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, v36
|
|
; 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, v34
|
|
; 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, v39
|
|
; 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 v21, v47
|
|
; 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 v22, v46
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; 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, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v44
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v41
|
|
; 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, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v40
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; 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, v33
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v12
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v11, v13, v11
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; 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: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; 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, v14, v12
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; 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: 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_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v15
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_or_b32_e32 v14, v16, v14
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v18
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_or_b32_e32 v17, v19, v17
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v45
|
|
; SI-NEXT: v_or_b32_e32 v20, v22, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v43
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v23, v25, v23
|
|
; 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 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v48f16_to_v24i32:
|
|
; 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: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_mov_b32_e32 v32, v23
|
|
; VI-NEXT: v_mov_b32_e32 v33, v22
|
|
; VI-NEXT: v_mov_b32_e32 v34, v21
|
|
; VI-NEXT: v_mov_b32_e32 v35, v20
|
|
; VI-NEXT: v_mov_b32_e32 v36, v19
|
|
; VI-NEXT: v_mov_b32_e32 v37, v18
|
|
; VI-NEXT: v_mov_b32_e32 v38, v17
|
|
; VI-NEXT: v_mov_b32_e32 v39, v16
|
|
; VI-NEXT: v_mov_b32_e32 v48, v15
|
|
; VI-NEXT: v_mov_b32_e32 v49, v14
|
|
; VI-NEXT: v_mov_b32_e32 v50, v13
|
|
; VI-NEXT: v_mov_b32_e32 v51, v12
|
|
; VI-NEXT: v_mov_b32_e32 v52, v11
|
|
; VI-NEXT: v_mov_b32_e32 v53, v10
|
|
; VI-NEXT: v_mov_b32_e32 v54, v9
|
|
; VI-NEXT: v_mov_b32_e32 v55, v8
|
|
; VI-NEXT: v_mov_b32_e32 v40, v7
|
|
; VI-NEXT: v_mov_b32_e32 v41, v6
|
|
; VI-NEXT: v_mov_b32_e32 v42, v5
|
|
; VI-NEXT: v_mov_b32_e32 v43, v4
|
|
; VI-NEXT: v_mov_b32_e32 v44, v3
|
|
; VI-NEXT: v_mov_b32_e32 v45, v2
|
|
; VI-NEXT: v_mov_b32_e32 v46, v1
|
|
; VI-NEXT: v_mov_b32_e32 v47, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 v23, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v23, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v23, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v23, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v23, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v23, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v23, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v23, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v23, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v23, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v23, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v23, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v23, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v23, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v23, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v23, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v23, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v23, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v23, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v23, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v23, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v23, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v23, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v23, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v44, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v43, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v42, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v41, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v40, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v55, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v53, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v52, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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 v23, 0x200
|
|
; VI-NEXT: v_add_f16_sdwa v0, v47, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v47
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_add_f16_sdwa v1, v46, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v2, 0x200, v46
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; VI-NEXT: v_add_f16_sdwa v2, v45, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v45
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v3, v44, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v44
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_add_f16_sdwa v4, v43, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v43
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_add_f16_sdwa v5, v42, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v42
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_add_f16_sdwa v6, v41, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v41
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_add_f16_sdwa v7, v40, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v40
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_add_f16_sdwa v8, v55, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v55
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_add_f16_sdwa v9, v54, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, 0x200, v54
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, 0x200, v53
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_add_f16_sdwa v11, v52, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, 0x200, v52
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_add_f16_sdwa v12, v51, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, 0x200, v51
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_add_f16_sdwa v13, v50, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v50
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_add_f16_sdwa v14, v49, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v48, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v39, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v38, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v37, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v36, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v35, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v34, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB18_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48f16_to_v24i32:
|
|
; 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 v33, v22
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v13
|
|
; 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, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v12
|
|
; 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, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v11
|
|
; 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, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v10
|
|
; 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, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v1
|
|
; 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, v53
|
|
; 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, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v47
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 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:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; 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: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; 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 v9, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: s_movk_i32 s7, 0x200
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, 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(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, 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: .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_v48f16_to_v24i32:
|
|
; 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, v24
|
|
; 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: .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_v48f16_to_v24i32:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v24
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v53, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v55, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v64, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v52, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v51, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v50, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v49, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v48, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v39, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v38, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v37, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v36, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v35, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v34, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v33, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v32, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v31, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v30, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v29, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v28, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v27, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v26, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v25, v23, 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: .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 <48 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <48 x half> %a1 to <24 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x half> %a to <24 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x i32> %phi
|
|
}
|
|
|
|
define inreg <24 x i32> @bitcast_v48f16_to_v24i32_scalar(<48 x half> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v48f16_to_v24i32_scalar:
|
|
; 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:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v62, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v63, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v61, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v60, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v59, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v58, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v57, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v56, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v45, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, s17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, s16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, s19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, s18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, s21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, s20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, s23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, s22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, s25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, s24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, s27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v30, s28
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v31
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v34
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v35
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, s26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, s29
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v40
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: s_cbranch_scc0 .LBB19_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v0, v8, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v7, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v6, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v5, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v34
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v62
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v52
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v48
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v60
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v58
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v46
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v4, v54, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v35, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v30, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v55, v7
|
|
; SI-NEXT: v_or_b32_e32 v8, v63, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v43, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v49, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v39, v11
|
|
; SI-NEXT: v_or_b32_e32 v12, v37, v12
|
|
; SI-NEXT: v_or_b32_e32 v13, v61, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v33, v14
|
|
; SI-NEXT: v_or_b32_e32 v15, v59, v15
|
|
; SI-NEXT: v_or_b32_e32 v16, v57, v16
|
|
; SI-NEXT: v_or_b32_e32 v17, v47, v17
|
|
; SI-NEXT: v_or_b32_e32 v18, v45, v18
|
|
; SI-NEXT: v_or_b32_e32 v19, v25, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v41, v20
|
|
; SI-NEXT: v_or_b32_e32 v21, v28, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v26, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v31, v23
|
|
; SI-NEXT: s_cbranch_execnz .LBB19_3
|
|
; SI-NEXT: .LBB19_2: ; %cmp.true
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v30
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v55
|
|
; 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_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; 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 v9, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v49
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v48
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v60
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v33
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; 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_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v59
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v56
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v47
|
|
; 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, v45
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v42
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v41
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v28
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, 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 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: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_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 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; 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, v51
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; 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, v34
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; 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, v62
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; 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, v53
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; 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, v52
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; 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, v50
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v38
|
|
; SI-NEXT: v_or_b32_e32 v11, v13, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v36
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v61
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v32
|
|
; SI-NEXT: v_or_b32_e32 v14, v16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v58
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v57
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v46
|
|
; SI-NEXT: v_or_b32_e32 v17, v19, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v44
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v31
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v29
|
|
; SI-NEXT: v_or_b32_e32 v20, v22, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v27
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v23, v25, v23
|
|
; SI-NEXT: .LBB19_3: ; %end
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB19_4:
|
|
; SI-NEXT: v_mov_b32_e32 v40, v31
|
|
; SI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v54, v33
|
|
; SI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v51, v32
|
|
; SI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v34, v59
|
|
; SI-NEXT: v_mov_b32_e32 v59, v46
|
|
; SI-NEXT: v_mov_b32_e32 v46, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v24
|
|
; SI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v35, v60
|
|
; SI-NEXT: v_mov_b32_e32 v60, v47
|
|
; SI-NEXT: v_mov_b32_e32 v47, v42
|
|
; SI-NEXT: v_mov_b32_e32 v42, v26
|
|
; SI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v55, v52
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_mov_b32_e32 v53, v43
|
|
; SI-NEXT: v_mov_b32_e32 v52, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v61
|
|
; SI-NEXT: v_mov_b32_e32 v61, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v25
|
|
; SI-NEXT: v_mov_b32_e32 v43, v27
|
|
; SI-NEXT: v_mov_b32_e32 v33, v62
|
|
; SI-NEXT: v_mov_b32_e32 v62, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v44
|
|
; SI-NEXT: v_mov_b32_e32 v44, v28
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v63, v58
|
|
; SI-NEXT: v_mov_b32_e32 v58, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v29
|
|
; SI-NEXT: v_mov_b32_e32 v32, v30
|
|
; 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: v_mov_b32_e32 v29, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v58
|
|
; SI-NEXT: v_mov_b32_e32 v58, v63
|
|
; SI-NEXT: v_mov_b32_e32 v27, v43
|
|
; SI-NEXT: v_mov_b32_e32 v25, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v61
|
|
; SI-NEXT: v_mov_b32_e32 v61, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v52
|
|
; SI-NEXT: v_mov_b32_e32 v43, v53
|
|
; SI-NEXT: v_mov_b32_e32 v52, v55
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v30, v32
|
|
; SI-NEXT: v_mov_b32_e32 v28, v44
|
|
; SI-NEXT: v_mov_b32_e32 v44, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v62
|
|
; SI-NEXT: v_mov_b32_e32 v62, v33
|
|
; SI-NEXT: v_mov_b32_e32 v26, v42
|
|
; SI-NEXT: v_mov_b32_e32 v42, v47
|
|
; SI-NEXT: v_mov_b32_e32 v47, v60
|
|
; SI-NEXT: v_mov_b32_e32 v60, v35
|
|
; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v24, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v46
|
|
; SI-NEXT: v_mov_b32_e32 v46, v59
|
|
; SI-NEXT: v_mov_b32_e32 v59, v34
|
|
; SI-NEXT: v_mov_b32_e32 v32, v51
|
|
; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v33, v54
|
|
; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v31, v40
|
|
; SI-NEXT: s_branch .LBB19_2
|
|
;
|
|
; VI-LABEL: bitcast_v48f16_to_v24i32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v32, v9
|
|
; VI-NEXT: v_mov_b32_e32 v33, v8
|
|
; VI-NEXT: v_mov_b32_e32 v34, v7
|
|
; VI-NEXT: v_mov_b32_e32 v35, v6
|
|
; VI-NEXT: v_mov_b32_e32 v36, v5
|
|
; VI-NEXT: v_mov_b32_e32 v37, v4
|
|
; VI-NEXT: v_mov_b32_e32 v38, v3
|
|
; VI-NEXT: v_mov_b32_e32 v39, v2
|
|
; VI-NEXT: v_mov_b32_e32 v48, v1
|
|
; VI-NEXT: v_mov_b32_e32 v49, v0
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: s_cbranch_scc0 .LBB19_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s43, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s44, s42, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s44
|
|
; VI-NEXT: s_and_b32 s44, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s45, s41, 16
|
|
; VI-NEXT: s_or_b32 s44, s44, s45
|
|
; VI-NEXT: s_and_b32 s45, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s46, s40, 16
|
|
; VI-NEXT: s_or_b32 s45, s45, s46
|
|
; VI-NEXT: s_and_b32 s46, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s47, s15, 16
|
|
; VI-NEXT: s_or_b32 s46, s46, s47
|
|
; VI-NEXT: s_and_b32 s47, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s56, s14, 16
|
|
; VI-NEXT: v_mov_b32_e32 v0, 16
|
|
; VI-NEXT: s_or_b32 s47, s47, s56
|
|
; VI-NEXT: s_and_b32 s56, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s57, s13, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s56, s56, s57
|
|
; VI-NEXT: s_and_b32 s57, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s58, s12, 16
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s57, s57, s58
|
|
; VI-NEXT: s_and_b32 s58, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s59, s11, 16
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s58, s58, s59
|
|
; VI-NEXT: s_and_b32 s59, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s60, s10, 16
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s59, s59, s60
|
|
; VI-NEXT: s_and_b32 s60, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s61, s9, 16
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s60, s60, s61
|
|
; VI-NEXT: s_and_b32 s61, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s62, s8, 16
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s61, s61, s62
|
|
; VI-NEXT: s_and_b32 s62, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s63, s7, 16
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s62, s62, s63
|
|
; VI-NEXT: s_and_b32 s63, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s72, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s63, s63, s72
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v0, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s44
|
|
; VI-NEXT: v_mov_b32_e32 v3, s45
|
|
; VI-NEXT: v_mov_b32_e32 v4, s46
|
|
; VI-NEXT: v_mov_b32_e32 v5, s47
|
|
; VI-NEXT: v_mov_b32_e32 v6, s56
|
|
; VI-NEXT: v_mov_b32_e32 v7, s57
|
|
; VI-NEXT: v_mov_b32_e32 v8, s58
|
|
; VI-NEXT: v_mov_b32_e32 v9, s59
|
|
; VI-NEXT: v_mov_b32_e32 v10, s60
|
|
; VI-NEXT: v_mov_b32_e32 v11, s61
|
|
; VI-NEXT: v_mov_b32_e32 v12, s62
|
|
; VI-NEXT: v_mov_b32_e32 v13, s63
|
|
; VI-NEXT: s_cbranch_execnz .LBB19_3
|
|
; VI-NEXT: .LBB19_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v13, 0x200
|
|
; VI-NEXT: v_mov_b32_e32 v0, s43
|
|
; VI-NEXT: v_add_f16_sdwa v0, v0, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, s16, v13
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_mov_b32_e32 v1, s42
|
|
; VI-NEXT: v_add_f16_sdwa v1, v1, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v2, s17, v13
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; VI-NEXT: v_mov_b32_e32 v2, s41
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s18, v13
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v3, s40
|
|
; VI-NEXT: v_add_f16_sdwa v3, v3, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, s19, v13
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_mov_b32_e32 v4, s15
|
|
; VI-NEXT: v_add_f16_sdwa v4, v4, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, s20, v13
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_mov_b32_e32 v5, s14
|
|
; VI-NEXT: v_add_f16_sdwa v5, v5, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, s21, v13
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_mov_b32_e32 v6, s13
|
|
; VI-NEXT: v_add_f16_sdwa v6, v6, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, s22, v13
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_mov_b32_e32 v7, s12
|
|
; VI-NEXT: v_add_f16_sdwa v7, v7, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, s23, v13
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_mov_b32_e32 v8, s11
|
|
; VI-NEXT: v_add_f16_sdwa v8, v8, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, s24, v13
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_mov_b32_e32 v9, s10
|
|
; VI-NEXT: v_add_f16_sdwa v9, v9, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, s25, v13
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_mov_b32_e32 v10, s9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v10, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, s26, v13
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_mov_b32_e32 v11, s8
|
|
; VI-NEXT: v_add_f16_sdwa v11, v11, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, s27, v13
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_mov_b32_e32 v12, s7
|
|
; VI-NEXT: v_add_f16_sdwa v12, v12, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, s28, v13
|
|
; VI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; VI-NEXT: v_mov_b32_e32 v14, s6
|
|
; VI-NEXT: v_add_f16_sdwa v14, v14, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, s29, v13
|
|
; VI-NEXT: v_mov_b32_e32 v23, 0x200
|
|
; VI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; VI-NEXT: v_add_f16_sdwa v14, v49, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v48, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v39, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v38, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v37, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v36, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v35, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v34, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB19_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB19_4:
|
|
; 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_branch .LBB19_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v48f16_to_v24i32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v0
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, 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: buffer_store_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v32
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v33
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v49
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s17, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s18, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s19, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s20, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s21, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s22, s12
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s23, s13
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s24, s14
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s25, s15
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s29, s40
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB19_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v33
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s19
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB19_3
|
|
; GFX9-NEXT: .LBB19_2: ; %cmp.true
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v49
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v48
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v39
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v38
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v37
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v36
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v35
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v34
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v33
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v32
|
|
; GFX9-NEXT: s_movk_i32 s4, 0x200
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, 0x200
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v23
|
|
; GFX9-NEXT: v_pk_add_f16 v0, s6, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, s7, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, s8, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, s9, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, s10, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, s11, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, s12, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, s13, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, s14, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v9, s15, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, s16, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, s17, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, s18, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, s19, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, v14, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, v15, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, v16, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, v17, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, v18, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, v19, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, v20, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, v21, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, v22, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, v23, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB19_3: ; %end
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB19_4:
|
|
; 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_branch .LBB19_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v24i32_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB19_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; 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, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB19_3
|
|
; GFX11-TRUE16-NEXT: .LBB19_2: ; %cmp.true
|
|
; 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, s4 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, s17 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: .LBB19_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB19_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB19_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v48f16_to_v24i32_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s4, s0, s4
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s44
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s16, s5
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s17, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s18, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s19, s8
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s20, s9
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s21, s10
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s22, s11
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s23, s12
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s24, s13
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s25, s14
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s26, s43
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB19_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB19_3
|
|
; GFX11-FAKE16-NEXT: .LBB19_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, s4 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 v2, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, s17 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: .LBB19_3: ; %end
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB19_4:
|
|
; GFX11-FAKE16-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
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB19_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <48 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <48 x half> %a1 to <24 x i32>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x half> %a to <24 x i32>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x i32> %phi
|
|
}
|
|
|
|
define <12 x i64> @bitcast_v24f32_to_v12i64(<24 x float> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v24f32_to_v12i64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24f32_to_v12i64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24f32_to_v12i64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24f32_to_v12i64:
|
|
; 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, v24
|
|
; 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 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 <24 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <24 x float> %a1 to <12 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x float> %a to <12 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x i64> %phi
|
|
}
|
|
|
|
define inreg <12 x i64> @bitcast_v24f32_to_v12i64_scalar(<24 x float> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v24f32_to_v12i64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: s_cbranch_scc0 .LBB21_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB21_3
|
|
; SI-NEXT: .LBB21_2: ; %cmp.true
|
|
; 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: .LBB21_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB21_4:
|
|
; SI-NEXT: s_branch .LBB21_2
|
|
;
|
|
; VI-LABEL: bitcast_v24f32_to_v12i64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: s_cbranch_scc0 .LBB21_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB21_3
|
|
; VI-NEXT: .LBB21_2: ; %cmp.true
|
|
; 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: .LBB21_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB21_4:
|
|
; VI-NEXT: s_branch .LBB21_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v24f32_to_v12i64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: s_cbranch_scc0 .LBB21_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB21_3
|
|
; GFX9-NEXT: .LBB21_2: ; %cmp.true
|
|
; 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: .LBB21_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB21_4:
|
|
; GFX9-NEXT: s_branch .LBB21_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v24f32_to_v12i64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; 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_vccz .LBB21_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB21_3:
|
|
; GFX11-NEXT: .LBB21_4: ; %cmp.true
|
|
; 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: 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 <24 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <24 x float> %a1 to <12 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x float> %a to <12 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x i64> %phi
|
|
}
|
|
|
|
define <24 x float> @bitcast_v12i64_to_v24f32(<12 x i64> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v12i64_to_v24f32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v12i64_to_v24f32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v12i64_to_v24f32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v12i64_to_v24f32:
|
|
; 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, v24
|
|
; 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 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 <12 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <12 x i64> %a1 to <24 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x i64> %a to <24 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x float> %phi
|
|
}
|
|
|
|
define inreg <24 x float> @bitcast_v12i64_to_v24f32_scalar(<12 x i64> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v12i64_to_v24f32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB23_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB23_4:
|
|
; SI-NEXT: s_branch .LBB23_2
|
|
;
|
|
; VI-LABEL: bitcast_v12i64_to_v24f32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB23_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB23_4:
|
|
; VI-NEXT: s_branch .LBB23_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v12i64_to_v24f32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB23_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB23_4:
|
|
; GFX9-NEXT: s_branch .LBB23_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v12i64_to_v24f32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB23_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccz .LBB23_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB23_3:
|
|
; GFX11-NEXT: .LBB23_4: ; %cmp.true
|
|
; 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: 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 <12 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <12 x i64> %a1 to <24 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x i64> %a to <24 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x float> %phi
|
|
}
|
|
|
|
define <12 x double> @bitcast_v24f32_to_v12f64(<24 x float> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v24f32_to_v12f64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24f32_to_v12f64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24f32_to_v12f64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 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_v24f32_to_v12f64:
|
|
; 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, v24
|
|
; 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 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 <24 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <24 x float> %a1 to <12 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x float> %a to <12 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x double> %phi
|
|
}
|
|
|
|
define inreg <12 x double> @bitcast_v24f32_to_v12f64_scalar(<24 x float> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v24f32_to_v12f64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: s_cbranch_scc0 .LBB25_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB25_3
|
|
; SI-NEXT: .LBB25_2: ; %cmp.true
|
|
; 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: .LBB25_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB25_4:
|
|
; SI-NEXT: s_branch .LBB25_2
|
|
;
|
|
; VI-LABEL: bitcast_v24f32_to_v12f64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: s_cbranch_scc0 .LBB25_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB25_3
|
|
; VI-NEXT: .LBB25_2: ; %cmp.true
|
|
; 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: .LBB25_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB25_4:
|
|
; VI-NEXT: s_branch .LBB25_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v24f32_to_v12f64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: s_cbranch_scc0 .LBB25_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB25_3
|
|
; GFX9-NEXT: .LBB25_2: ; %cmp.true
|
|
; 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: .LBB25_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB25_4:
|
|
; GFX9-NEXT: s_branch .LBB25_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v24f32_to_v12f64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; 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_vccz .LBB25_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB25_3:
|
|
; GFX11-NEXT: .LBB25_4: ; %cmp.true
|
|
; 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: 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 <24 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <24 x float> %a1 to <12 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x float> %a to <12 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x double> %phi
|
|
}
|
|
|
|
define <24 x float> @bitcast_v12f64_to_v24f32(<12 x double> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v12f64_to_v24f32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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[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_v12f64_to_v24f32:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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[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_v12f64_to_v24f32:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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[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_v12f64_to_v24f32:
|
|
; 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, v24
|
|
; 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[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 <12 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <12 x double> %a1 to <24 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x double> %a to <24 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x float> %phi
|
|
}
|
|
|
|
define inreg <24 x float> @bitcast_v12f64_to_v24f32_scalar(<12 x double> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v12f64_to_v24f32_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: s_cbranch_scc0 .LBB27_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB27_3
|
|
; SI-NEXT: .LBB27_2: ; %cmp.true
|
|
; 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: .LBB27_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB27_4:
|
|
; SI-NEXT: s_branch .LBB27_2
|
|
;
|
|
; VI-LABEL: bitcast_v12f64_to_v24f32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: s_cbranch_scc0 .LBB27_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB27_3
|
|
; VI-NEXT: .LBB27_2: ; %cmp.true
|
|
; 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: .LBB27_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB27_4:
|
|
; VI-NEXT: s_branch .LBB27_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v12f64_to_v24f32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB27_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB27_3
|
|
; GFX9-NEXT: .LBB27_2: ; %cmp.true
|
|
; 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: .LBB27_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB27_4:
|
|
; GFX9-NEXT: s_branch .LBB27_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v12f64_to_v24f32_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; 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_vccz .LBB27_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB27_3:
|
|
; GFX11-NEXT: .LBB27_4: ; %cmp.true
|
|
; 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: 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 <12 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <12 x double> %a1 to <24 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x double> %a to <24 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x float> %phi
|
|
}
|
|
|
|
define <48 x i16> @bitcast_v24f32_to_v48i16(<24 x float> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v24f32_to_v48i16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v25
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; 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 v25, v24, v23, 16
|
|
; SI-NEXT: v_alignbit_b32 v26, v22, v21, 16
|
|
; SI-NEXT: v_alignbit_b32 v27, v20, v19, 16
|
|
; SI-NEXT: v_alignbit_b32 v28, v18, v17, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v16, v15, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v14, v13, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v12, v11, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v10, v9, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v8, v7, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v6, v5, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v4, v3, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v2, v1, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v2
|
|
; 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 v2, 1.0, v2
|
|
; SI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; 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 v6, 1.0, v6
|
|
; SI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; 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 v10, 1.0, v10
|
|
; SI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; 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 v14, 1.0, v14
|
|
; SI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; 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 v18, 1.0, v18
|
|
; SI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; 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 v22, 1.0, v22
|
|
; SI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; SI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; SI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; SI-NEXT: v_alignbit_b32 v25, v24, v23, 16
|
|
; SI-NEXT: v_alignbit_b32 v26, v22, v21, 16
|
|
; SI-NEXT: v_alignbit_b32 v27, v20, v19, 16
|
|
; SI-NEXT: v_alignbit_b32 v28, v18, v17, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v16, v15, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v14, v13, 16
|
|
; SI-NEXT: v_alignbit_b32 v34, v12, v11, 16
|
|
; SI-NEXT: v_alignbit_b32 v36, v10, v9, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v8, v7, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v6, v5, 16
|
|
; SI-NEXT: v_alignbit_b32 v51, v4, v3, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v2, v1, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v2
|
|
; SI-NEXT: .LBB28_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v53
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 4, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 8, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 12, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 16, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 20, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 24, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 28, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 32, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 36, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 40, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v31
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 52, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v29
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 56, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 60, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 64, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v27
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x4c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x50, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x54, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x58, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v30
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v24f32_to_v48i16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; VI-NEXT: .LBB28_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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_lshlrev_b32_e32 v27, 16, v27
|
|
; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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_or_b32_sdwa v20, v20, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v24f32_to_v48i16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v54, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v53, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v52, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v51, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v50, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v49, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v48, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v39, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v37, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v36, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v35, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v34, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v33, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v32, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v31, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v30, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v29, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v28, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v27, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v26, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v25, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s4
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v24f32_to_v48i16:
|
|
; 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, v24
|
|
; 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 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_v24f32_to_v48i16:
|
|
; 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, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v53, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v52, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v51, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v50, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v49, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v48, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v39, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v37, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v35, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v33, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v32, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v31, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v30, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v29, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v28, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v27, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v26, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v25, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v24, v23, 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 <24 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <24 x float> %a1 to <48 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x float> %a to <48 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x i16> %phi
|
|
}
|
|
|
|
define inreg <48 x i16> @bitcast_v24f32_to_v48i16_scalar(<24 x float> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v24f32_to_v48i16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
|
|
; SI-NEXT: v_mov_b32_e32 v23, s16
|
|
; SI-NEXT: v_mov_b32_e32 v24, s17
|
|
; SI-NEXT: v_mov_b32_e32 v21, s18
|
|
; SI-NEXT: v_mov_b32_e32 v22, s19
|
|
; SI-NEXT: v_mov_b32_e32 v19, s20
|
|
; SI-NEXT: v_mov_b32_e32 v20, s21
|
|
; SI-NEXT: v_mov_b32_e32 v17, s22
|
|
; SI-NEXT: v_mov_b32_e32 v18, s23
|
|
; SI-NEXT: v_mov_b32_e32 v15, s24
|
|
; SI-NEXT: v_mov_b32_e32 v16, s25
|
|
; SI-NEXT: v_mov_b32_e32 v13, s26
|
|
; SI-NEXT: v_mov_b32_e32 v14, s27
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: v_mov_b32_e32 v11, s28
|
|
; SI-NEXT: v_mov_b32_e32 v12, s29
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_cbranch_scc0 .LBB29_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_lshr_b64 v[25:26], v[9:10], 16
|
|
; SI-NEXT: v_lshr_b64 v[32:33], v[11:12], 16
|
|
; SI-NEXT: v_lshr_b64 v[26:27], v[7:8], 16
|
|
; SI-NEXT: v_lshr_b64 v[33:34], v[15:16], 16
|
|
; SI-NEXT: v_lshr_b64 v[27:28], v[5:6], 16
|
|
; SI-NEXT: v_lshr_b64 v[34:35], v[17:18], 16
|
|
; SI-NEXT: v_lshr_b64 v[28:29], v[3:4], 16
|
|
; SI-NEXT: v_lshr_b64 v[35:36], v[19:20], 16
|
|
; SI-NEXT: v_lshr_b64 v[29:30], v[1:2], 16
|
|
; SI-NEXT: v_lshr_b64 v[36:37], v[21:22], 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v20
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v22
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v24
|
|
; SI-NEXT: v_lshr_b64 v[30:31], v[13:14], 16
|
|
; SI-NEXT: v_lshr_b64 v[37:38], v[23:24], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB29_3
|
|
; SI-NEXT: .LBB29_2: ; %cmp.true
|
|
; 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 v16, 1.0, v16
|
|
; SI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; SI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; SI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; SI-NEXT: v_lshr_b64 v[25:26], v[9:10], 16
|
|
; SI-NEXT: v_lshr_b64 v[32:33], v[11:12], 16
|
|
; 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 v6, 1.0, v6
|
|
; SI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; SI-NEXT: v_lshr_b64 v[26:27], v[7:8], 16
|
|
; SI-NEXT: v_lshr_b64 v[33:34], v[15:16], 16
|
|
; 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 v4, 1.0, v4
|
|
; SI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; SI-NEXT: v_lshr_b64 v[27:28], v[5:6], 16
|
|
; SI-NEXT: v_lshr_b64 v[34:35], v[17:18], 16
|
|
; 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 v2, 1.0, v2
|
|
; SI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; SI-NEXT: v_lshr_b64 v[28:29], v[3:4], 16
|
|
; SI-NEXT: v_lshr_b64 v[35:36], v[19:20], 16
|
|
; 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 v14, 1.0, v14
|
|
; SI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; SI-NEXT: v_lshr_b64 v[29:30], v[1:2], 16
|
|
; SI-NEXT: v_lshr_b64 v[36:37], v[21:22], 16
|
|
; SI-NEXT: v_lshr_b64 v[30:31], v[13:14], 16
|
|
; SI-NEXT: v_lshr_b64 v[37:38], v[23:24], 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v20
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v22
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v24
|
|
; SI-NEXT: .LBB29_3: ; %end
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v37
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v31
|
|
; SI-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 4, v0
|
|
; SI-NEXT: buffer_store_dword v23, v24, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v23
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 8, v0
|
|
; SI-NEXT: buffer_store_dword v21, v23, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v41
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 12, v0
|
|
; SI-NEXT: buffer_store_dword v21, v22, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v21
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 16, v0
|
|
; SI-NEXT: buffer_store_dword v19, v21, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 20, v0
|
|
; SI-NEXT: buffer_store_dword v19, v20, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v19
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 24, v0
|
|
; SI-NEXT: buffer_store_dword v17, v19, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 28, v0
|
|
; SI-NEXT: buffer_store_dword v17, v18, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v17
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 32, v0
|
|
; SI-NEXT: buffer_store_dword v15, v17, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 36, v0
|
|
; SI-NEXT: buffer_store_dword v15, v16, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v30
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v15
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 40, v0
|
|
; SI-NEXT: buffer_store_dword v13, v15, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 44, v0
|
|
; SI-NEXT: buffer_store_dword v13, v14, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v13
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 48, v0
|
|
; SI-NEXT: buffer_store_dword v11, v13, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 52, v0
|
|
; SI-NEXT: buffer_store_dword v11, v12, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v29
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v11
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 56, v0
|
|
; SI-NEXT: buffer_store_dword v1, v11, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 60, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 64, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v27
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x4c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x50, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x54, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x58, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB29_4:
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; 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: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: s_branch .LBB29_2
|
|
;
|
|
; VI-LABEL: bitcast_v24f32_to_v48i16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, s16
|
|
; VI-NEXT: v_mov_b32_e32 v22, s17
|
|
; VI-NEXT: v_mov_b32_e32 v21, s18
|
|
; VI-NEXT: v_mov_b32_e32 v19, s19
|
|
; VI-NEXT: v_mov_b32_e32 v17, s20
|
|
; VI-NEXT: v_mov_b32_e32 v15, s21
|
|
; VI-NEXT: v_mov_b32_e32 v13, s22
|
|
; VI-NEXT: v_mov_b32_e32 v12, s23
|
|
; VI-NEXT: v_mov_b32_e32 v11, s24
|
|
; VI-NEXT: v_mov_b32_e32 v10, s25
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_mov_b32_e32 v20, s26
|
|
; VI-NEXT: v_mov_b32_e32 v18, s27
|
|
; VI-NEXT: v_mov_b32_e32 v16, s28
|
|
; VI-NEXT: v_mov_b32_e32 v14, s29
|
|
; VI-NEXT: s_cbranch_scc0 .LBB29_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; VI-NEXT: s_cbranch_execnz .LBB29_3
|
|
; VI-NEXT: .LBB29_2: ; %cmp.true
|
|
; 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_add_f32_e32 v14, 1.0, v14
|
|
; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; VI-NEXT: .LBB29_3: ; %end
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_or_b32_sdwa v30, v23, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v31
|
|
; VI-NEXT: v_or_b32_sdwa v31, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v32
|
|
; VI-NEXT: v_or_b32_sdwa v32, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v33
|
|
; VI-NEXT: v_or_b32_sdwa v33, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v24, v17, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v25
|
|
; VI-NEXT: v_or_b32_sdwa v25, v15, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v15, 16, v26
|
|
; VI-NEXT: v_or_b32_sdwa v26, v13, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v27
|
|
; VI-NEXT: v_or_b32_sdwa v27, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v52
|
|
; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v14, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v14, v0, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v50
|
|
; VI-NEXT: v_or_b32_sdwa v28, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v53
|
|
; VI-NEXT: v_or_b32_sdwa v15, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v29
|
|
; VI-NEXT: v_or_b32_sdwa v12, v16, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v48
|
|
; VI-NEXT: v_or_b32_sdwa v29, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v54
|
|
; VI-NEXT: v_or_b32_sdwa v17, v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v39
|
|
; VI-NEXT: v_or_b32_sdwa v11, v18, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v10, 16, v55
|
|
; VI-NEXT: v_or_b32_sdwa v19, v5, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v37
|
|
; VI-NEXT: v_or_b32_sdwa v10, v20, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v6, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v36
|
|
; VI-NEXT: v_or_b32_sdwa v21, v7, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v35
|
|
; VI-NEXT: v_or_b32_sdwa v22, v8, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v34
|
|
; VI-NEXT: v_or_b32_sdwa v23, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, v30
|
|
; VI-NEXT: v_mov_b32_e32 v1, v31
|
|
; VI-NEXT: v_mov_b32_e32 v2, v32
|
|
; VI-NEXT: v_mov_b32_e32 v3, v33
|
|
; VI-NEXT: v_mov_b32_e32 v4, v24
|
|
; VI-NEXT: v_mov_b32_e32 v5, v25
|
|
; VI-NEXT: v_mov_b32_e32 v6, v26
|
|
; VI-NEXT: v_mov_b32_e32 v7, v27
|
|
; VI-NEXT: v_mov_b32_e32 v8, v28
|
|
; VI-NEXT: v_mov_b32_e32 v9, v29
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB29_4:
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr24
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; 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: s_branch .LBB29_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v24f32_to_v48i16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s25
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s29
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB29_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB29_3
|
|
; GFX9-NEXT: .LBB29_2: ; %cmp.true
|
|
; 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_add_f32_e32 v14, 1.0, v14
|
|
; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; GFX9-NEXT: .LBB29_3: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v26, 16, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v25, 16, v15
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v50, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v2
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v27, 16, v12
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v49, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v24, 16, v17
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v48, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v4
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v28, v28, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v39, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v33, v33, 16, v19
|
|
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v38, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v6
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v29, v29, 16, v10
|
|
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v37, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v7
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v32, v32, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v36, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v30, v30, 16, v23
|
|
; GFX9-NEXT: v_lshl_or_b32 v31, v31, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v35, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v55, 16, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v54, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v53, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v52, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v34, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, v30
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, v31
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, v32
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, v25
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, v26
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, v27
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, v28
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, v29
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB29_4:
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; 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: s_branch .LBB29_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v24f32_to_v48i16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB29_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB29_3
|
|
; GFX11-TRUE16-NEXT: .LBB29_2: ; %cmp.true
|
|
; 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: v_lshrrev_b32_e32 v24, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; GFX11-TRUE16-NEXT: .LBB29_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v55, v55 :: v_dual_mov_b32 v54, v54
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v53 :: v_dual_mov_b32 v52, v52
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v51, v51 :: v_dual_mov_b32 v50, v50
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v49, v49 :: v_dual_mov_b32 v48, v48
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, v39 :: v_dual_mov_b32 v38, v38
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, v37 :: v_dual_mov_b32 v36, v36
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, v35 :: v_dual_mov_b32 v34, v34
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, v33 :: v_dual_mov_b32 v32, v32
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, v31 :: v_dual_mov_b32 v30, v30
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v29 :: v_dual_mov_b32 v28, v28
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v27 :: v_dual_mov_b32 v26, v26
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v25 :: v_dual_mov_b32 v24, v24
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v31.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v28.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v27.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v26.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v25.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v24.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB29_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB29_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v24f32_to_v48i16_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v23, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v21, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v19, s17
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v7, s20
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s19 :: v_dual_mov_b32 v13, s21
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v9, s25
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v14, s27
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB29_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v6, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB29_3
|
|
; GFX11-FAKE16-NEXT: .LBB29_2: ; %cmp.true
|
|
; 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_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v8, 1.0, v8
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v18, 1.0, v18
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v6, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: .LBB29_3: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v6, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v49, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v29, v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v39, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v38, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v52, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v48, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v16
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v28, v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v33, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v54, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v53, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v37, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v34, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v32, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v31, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v30, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v0, v24
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v28
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB29_4:
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr6
|
|
; 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: s_branch .LBB29_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <24 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <24 x float> %a1 to <48 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x float> %a to <48 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x i16> %phi
|
|
}
|
|
|
|
define <24 x float> @bitcast_v48i16_to_v24f32(<48 x i16> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v48i16_to_v24f32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v48, v14
|
|
; SI-NEXT: v_mov_b32_e32 v49, v12
|
|
; SI-NEXT: v_mov_b32_e32 v50, v10
|
|
; SI-NEXT: v_mov_b32_e32 v51, v8
|
|
; SI-NEXT: v_mov_b32_e32 v52, v6
|
|
; SI-NEXT: v_mov_b32_e32 v53, v4
|
|
; SI-NEXT: v_mov_b32_e32 v54, v2
|
|
; SI-NEXT: v_mov_b32_e32 v55, v0
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:68
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:60
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshlrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v43, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v42, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v41, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v29
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v8
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v12
|
|
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:52
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:44
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:40
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:36
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:196 ; 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 v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v55
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v54
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v53
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v52
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v51
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v50
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v49
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v48
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v47
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v39
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v46
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v38
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v45
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v37
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v44
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v36
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; 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: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v43
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v35
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v42
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v34
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v41
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v33
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v40
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v32
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v63
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v62
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v61
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v60
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v59
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v58
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v57
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v56
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; 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 v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v55
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v54
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v53
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v52
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v51
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v50
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v49
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v48
|
|
; 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_or_b32_e32 v0, v47, v0
|
|
; SI-NEXT: s_mov_b32 s6, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v1, v39, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v46, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v38, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v45, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v37, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v44, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v36, v7
|
|
; 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: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; 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_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v8, v43, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v35, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v42, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v34, v11
|
|
; SI-NEXT: v_or_b32_e32 v12, v41, v12
|
|
; SI-NEXT: v_or_b32_e32 v13, v33, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v40, v14
|
|
; SI-NEXT: v_or_b32_e32 v15, v32, v15
|
|
; SI-NEXT: v_or_b32_e32 v16, v63, v16
|
|
; SI-NEXT: v_or_b32_e32 v17, v62, v17
|
|
; SI-NEXT: v_or_b32_e32 v18, v61, v18
|
|
; SI-NEXT: v_or_b32_e32 v19, v60, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v59, v20
|
|
; SI-NEXT: v_or_b32_e32 v21, v58, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v57, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v56, v23
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9
|
|
; 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: v_add_i32_e32 v20, vcc, s6, v20
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v22
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 0x30000, v23
|
|
; 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 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v48i16_to_v24f32:
|
|
; 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: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_mov_b32_e32 v32, v23
|
|
; VI-NEXT: v_mov_b32_e32 v33, v22
|
|
; VI-NEXT: v_mov_b32_e32 v34, v21
|
|
; VI-NEXT: v_mov_b32_e32 v35, v20
|
|
; VI-NEXT: v_mov_b32_e32 v36, v19
|
|
; VI-NEXT: v_mov_b32_e32 v37, v18
|
|
; VI-NEXT: v_mov_b32_e32 v38, v17
|
|
; VI-NEXT: v_mov_b32_e32 v39, v16
|
|
; VI-NEXT: v_mov_b32_e32 v48, v15
|
|
; VI-NEXT: v_mov_b32_e32 v49, v14
|
|
; VI-NEXT: v_mov_b32_e32 v50, v13
|
|
; VI-NEXT: v_mov_b32_e32 v51, v12
|
|
; VI-NEXT: v_mov_b32_e32 v52, v11
|
|
; VI-NEXT: v_mov_b32_e32 v53, v10
|
|
; VI-NEXT: v_mov_b32_e32 v54, v9
|
|
; VI-NEXT: v_mov_b32_e32 v55, v8
|
|
; VI-NEXT: v_mov_b32_e32 v40, v7
|
|
; VI-NEXT: v_mov_b32_e32 v41, v6
|
|
; VI-NEXT: v_mov_b32_e32 v42, v5
|
|
; VI-NEXT: v_mov_b32_e32 v43, v4
|
|
; VI-NEXT: v_mov_b32_e32 v44, v3
|
|
; VI-NEXT: v_mov_b32_e32 v45, v2
|
|
; VI-NEXT: v_mov_b32_e32 v46, v1
|
|
; VI-NEXT: v_mov_b32_e32 v47, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 v23, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v23, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v23, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v23, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v23, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v23, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v23, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v23, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v23, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v23, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v23, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v23, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v23, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v23, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v23, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v23, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v23, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v23, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v23, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v23, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v23, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v23, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v23, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v23, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v44, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v43, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v42, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v41, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v40, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v55, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v53, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v52, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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 v23, 3
|
|
; VI-NEXT: v_add_u16_e32 v0, 3, v47
|
|
; VI-NEXT: v_add_u16_sdwa v1, v47, v23 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_add_u16_e32 v1, 3, v46
|
|
; VI-NEXT: v_add_u16_sdwa v2, v46, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v45
|
|
; VI-NEXT: v_add_u16_sdwa v3, v45, v23 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, v44
|
|
; VI-NEXT: v_add_u16_sdwa v4, v44, v23 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, v43
|
|
; VI-NEXT: v_add_u16_sdwa v5, v43, v23 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, v42
|
|
; VI-NEXT: v_add_u16_sdwa v6, v42, v23 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, v41
|
|
; VI-NEXT: v_add_u16_sdwa v7, v41, v23 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, v40
|
|
; VI-NEXT: v_add_u16_sdwa v8, v40, v23 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, v55
|
|
; VI-NEXT: v_add_u16_sdwa v9, v55, v23 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, v54
|
|
; VI-NEXT: v_add_u16_sdwa v10, v54, v23 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, v53
|
|
; VI-NEXT: v_add_u16_sdwa v11, v53, v23 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, v52
|
|
; VI-NEXT: v_add_u16_sdwa v12, v52, v23 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, v51
|
|
; VI-NEXT: v_add_u16_sdwa v13, v51, v23 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, v50
|
|
; VI-NEXT: v_add_u16_sdwa v14, v50, v23 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, v49
|
|
; VI-NEXT: v_add_u16_sdwa v15, v49, v23 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, v48
|
|
; VI-NEXT: v_add_u16_sdwa v16, v48, v23 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, v39
|
|
; VI-NEXT: v_add_u16_sdwa v17, v39, v23 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, v38
|
|
; VI-NEXT: v_add_u16_sdwa v18, v38, v23 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, v37
|
|
; VI-NEXT: v_add_u16_sdwa v19, v37, v23 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, v36
|
|
; VI-NEXT: v_add_u16_sdwa v20, v36, v23 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, v35
|
|
; VI-NEXT: v_add_u16_sdwa v21, v35, v23 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, v34
|
|
; VI-NEXT: v_add_u16_sdwa v22, v34, v23 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, v33
|
|
; VI-NEXT: v_add_u16_sdwa v24, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v22, v22, v24
|
|
; VI-NEXT: v_add_u16_e32 v24, 3, v32
|
|
; VI-NEXT: v_add_u16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB30_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48i16_to_v24f32:
|
|
; 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 v33, v22
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v13
|
|
; 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, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v12
|
|
; 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, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v11
|
|
; 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, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v10
|
|
; 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, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v1
|
|
; 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, v53
|
|
; 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, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v47
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 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:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; 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: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; 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 v9, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, 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(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, 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: .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_v48i16_to_v24f32:
|
|
; 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, v24
|
|
; 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: .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_v48i16_to_v24f32:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v24
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v53, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v55, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v64, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v52, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v51, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v50, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v49, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v48, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v39, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v38, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v37, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v36, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v35, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v34, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v33, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v32, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v31, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v30, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v29, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v28, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v27, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v26, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v25, v23, 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: .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 <48 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <48 x i16> %a1 to <24 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x i16> %a to <24 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x float> %phi
|
|
}
|
|
|
|
define inreg <24 x float> @bitcast_v48i16_to_v24f32_scalar(<48 x i16> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v48i16_to_v24f32_scalar:
|
|
; 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:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v61, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v63, v2
|
|
; SI-NEXT: v_mov_b32_e32 v56, v0
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: v_mov_b32_e32 v31, v22
|
|
; SI-NEXT: v_mov_b32_e32 v34, v20
|
|
; SI-NEXT: v_mov_b32_e32 v35, v18
|
|
; SI-NEXT: v_mov_b32_e32 v36, v16
|
|
; SI-NEXT: v_mov_b32_e32 v37, v14
|
|
; SI-NEXT: v_mov_b32_e32 v38, v12
|
|
; SI-NEXT: v_mov_b32_e32 v39, v10
|
|
; SI-NEXT: v_mov_b32_e32 v48, v8
|
|
; SI-NEXT: v_mov_b32_e32 v50, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v55, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v41, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v52, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v2
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v4
|
|
; SI-NEXT: s_cbranch_scc0 .LBB31_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v56
|
|
; SI-NEXT: v_or_b32_e32 v7, v0, v54
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v61
|
|
; SI-NEXT: v_or_b32_e32 v9, v0, v32
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v50
|
|
; SI-NEXT: v_or_b32_e32 v10, v0, v55
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; SI-NEXT: v_or_b32_e32 v11, v0, v62
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; SI-NEXT: v_or_b32_e32 v12, v0, v46
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; SI-NEXT: v_or_b32_e32 v13, v0, v45
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; SI-NEXT: v_or_b32_e32 v14, v0, v60
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; SI-NEXT: v_or_b32_e32 v15, v0, v59
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s17, 16
|
|
; SI-NEXT: v_or_b32_e32 v16, v0, v58
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; SI-NEXT: s_or_b32 s4, s4, s5
|
|
; SI-NEXT: s_and_b32 s5, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s19, 16
|
|
; SI-NEXT: v_or_b32_e32 v17, v0, v41
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v31
|
|
; SI-NEXT: s_or_b32 s5, s5, s6
|
|
; SI-NEXT: s_and_b32 s6, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s21, 16
|
|
; SI-NEXT: v_or_b32_e32 v18, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v24
|
|
; SI-NEXT: s_or_b32 s6, s6, s7
|
|
; SI-NEXT: s_and_b32 s7, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s23, 16
|
|
; SI-NEXT: v_or_b32_e32 v19, v0, v52
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v26
|
|
; SI-NEXT: s_or_b32 s7, s7, s8
|
|
; SI-NEXT: s_and_b32 s8, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s25, 16
|
|
; SI-NEXT: v_or_b32_e32 v20, v0, v51
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v28
|
|
; SI-NEXT: s_or_b32 s8, s8, s9
|
|
; SI-NEXT: s_and_b32 s9, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s10, s27, 16
|
|
; SI-NEXT: v_or_b32_e32 v21, v0, v29
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v30
|
|
; SI-NEXT: s_or_b32 s9, s9, s10
|
|
; SI-NEXT: s_and_b32 s10, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s11, s29, 16
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v63
|
|
; SI-NEXT: v_or_b32_e32 v22, v0, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; SI-NEXT: s_or_b32 s10, s10, s11
|
|
; SI-NEXT: v_or_b32_e32 v8, v1, v57
|
|
; SI-NEXT: v_or_b32_e32 v23, v0, v25
|
|
; SI-NEXT: v_mov_b32_e32 v0, s4
|
|
; SI-NEXT: v_mov_b32_e32 v1, s5
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: v_mov_b32_e32 v3, s7
|
|
; SI-NEXT: v_mov_b32_e32 v4, s8
|
|
; SI-NEXT: v_mov_b32_e32 v5, s9
|
|
; SI-NEXT: v_mov_b32_e32 v6, s10
|
|
; SI-NEXT: s_cbranch_execnz .LBB31_3
|
|
; SI-NEXT: .LBB31_2: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v56
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v54, v0
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v61
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v32, v0
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v50
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_or_b32_e32 v0, v55, v0
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v48
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v62, v0
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v39
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v0, v46, v0
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v38
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v45, v0
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v37
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v60, v0
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v36
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v59, v0
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v35
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v58, v0
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v34
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v41, v0
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v31
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v40, v0
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v24
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v52, v0
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v26
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v51, v0
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s17, 16
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v28
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_and_b32 s5, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s19, 16
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_or_b32 s5, s6, s5
|
|
; SI-NEXT: s_and_b32 s6, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s21, 16
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v29, v0
|
|
; SI-NEXT: s_or_b32 s6, s7, s6
|
|
; SI-NEXT: s_and_b32 s7, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s23, 16
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v30
|
|
; SI-NEXT: s_or_b32 s7, s8, s7
|
|
; SI-NEXT: s_and_b32 s8, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s25, 16
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_or_b32 s8, s9, s8
|
|
; SI-NEXT: s_and_b32 s9, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s10, s27, 16
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v27, v0
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v63
|
|
; SI-NEXT: s_or_b32 s9, s10, s9
|
|
; SI-NEXT: s_and_b32 s10, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s11, s29, 16
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v49
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: s_or_b32 s10, s11, s10
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v57, v1
|
|
; SI-NEXT: s_add_i32 s4, s4, 0x30000
|
|
; SI-NEXT: s_add_i32 s5, s5, 0x30000
|
|
; SI-NEXT: s_add_i32 s6, s6, 0x30000
|
|
; SI-NEXT: s_add_i32 s7, s7, 0x30000
|
|
; SI-NEXT: s_add_i32 s8, s8, 0x30000
|
|
; SI-NEXT: s_add_i32 s9, s9, 0x30000
|
|
; SI-NEXT: s_add_i32 s10, s10, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v0, v25, v0
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 0x30000, v1
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 0x30000, v0
|
|
; SI-NEXT: v_mov_b32_e32 v0, s4
|
|
; SI-NEXT: v_mov_b32_e32 v1, s5
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: v_mov_b32_e32 v3, s7
|
|
; SI-NEXT: v_mov_b32_e32 v4, s8
|
|
; SI-NEXT: v_mov_b32_e32 v5, s9
|
|
; SI-NEXT: v_mov_b32_e32 v6, s10
|
|
; SI-NEXT: .LBB31_3: ; %end
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB31_4:
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v46, v51
|
|
; SI-NEXT: v_mov_b32_e32 v51, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v34
|
|
; SI-NEXT: v_mov_b32_e32 v34, v30
|
|
; SI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_mov_b32_e32 v47, v52
|
|
; SI-NEXT: v_mov_b32_e32 v52, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v35
|
|
; SI-NEXT: v_mov_b32_e32 v35, v28
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v55, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v41
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v41, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v26
|
|
; SI-NEXT: v_mov_b32_e32 v42, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v24
|
|
; SI-NEXT: v_mov_b32_e32 v33, v32
|
|
; SI-NEXT: v_mov_b32_e32 v32, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v40
|
|
; SI-NEXT: v_mov_b32_e32 v40, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v31
|
|
; SI-NEXT: v_mov_b32_e32 v43, v25
|
|
; SI-NEXT: v_mov_b32_e32 v44, v27
|
|
; SI-NEXT: v_mov_b32_e32 v53, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v29
|
|
; 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: v_mov_b32_e32 v26, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v55
|
|
; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v30, v34
|
|
; SI-NEXT: v_mov_b32_e32 v34, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v51
|
|
; SI-NEXT: v_mov_b32_e32 v51, v46
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v29, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v53
|
|
; SI-NEXT: v_mov_b32_e32 v27, v44
|
|
; SI-NEXT: v_mov_b32_e32 v25, v43
|
|
; SI-NEXT: v_mov_b32_e32 v31, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v40
|
|
; SI-NEXT: v_mov_b32_e32 v40, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v32
|
|
; SI-NEXT: v_mov_b32_e32 v32, v33
|
|
; SI-NEXT: v_mov_b32_e32 v24, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v42
|
|
; SI-NEXT: v_mov_b32_e32 v28, v35
|
|
; SI-NEXT: v_mov_b32_e32 v35, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v52
|
|
; SI-NEXT: v_mov_b32_e32 v52, v47
|
|
; SI-NEXT: s_branch .LBB31_2
|
|
;
|
|
; VI-LABEL: bitcast_v48i16_to_v24f32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v32, v9
|
|
; VI-NEXT: v_mov_b32_e32 v33, v8
|
|
; VI-NEXT: v_mov_b32_e32 v34, v7
|
|
; VI-NEXT: v_mov_b32_e32 v35, v6
|
|
; VI-NEXT: v_mov_b32_e32 v36, v5
|
|
; VI-NEXT: v_mov_b32_e32 v37, v4
|
|
; VI-NEXT: v_mov_b32_e32 v38, v3
|
|
; VI-NEXT: v_mov_b32_e32 v39, v2
|
|
; VI-NEXT: v_mov_b32_e32 v48, v1
|
|
; VI-NEXT: v_mov_b32_e32 v49, v0
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; 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, s43, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s44, s42, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s44
|
|
; VI-NEXT: s_and_b32 s44, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s45, s41, 16
|
|
; VI-NEXT: s_or_b32 s44, s44, s45
|
|
; VI-NEXT: s_and_b32 s45, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s46, s40, 16
|
|
; VI-NEXT: s_or_b32 s45, s45, s46
|
|
; VI-NEXT: s_and_b32 s46, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s47, s15, 16
|
|
; VI-NEXT: s_or_b32 s46, s46, s47
|
|
; VI-NEXT: s_and_b32 s47, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s56, s14, 16
|
|
; VI-NEXT: v_mov_b32_e32 v0, 16
|
|
; VI-NEXT: s_or_b32 s47, s47, s56
|
|
; VI-NEXT: s_and_b32 s56, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s57, s13, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s56, s56, s57
|
|
; VI-NEXT: s_and_b32 s57, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s58, s12, 16
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s57, s57, s58
|
|
; VI-NEXT: s_and_b32 s58, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s59, s11, 16
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s58, s58, s59
|
|
; VI-NEXT: s_and_b32 s59, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s60, s10, 16
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s59, s59, s60
|
|
; VI-NEXT: s_and_b32 s60, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s61, s9, 16
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s60, s60, s61
|
|
; VI-NEXT: s_and_b32 s61, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s62, s8, 16
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s61, s61, s62
|
|
; VI-NEXT: s_and_b32 s62, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s63, s7, 16
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s62, s62, s63
|
|
; VI-NEXT: s_and_b32 s63, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s72, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s63, s63, s72
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v0, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s44
|
|
; VI-NEXT: v_mov_b32_e32 v3, s45
|
|
; VI-NEXT: v_mov_b32_e32 v4, s46
|
|
; VI-NEXT: v_mov_b32_e32 v5, s47
|
|
; VI-NEXT: v_mov_b32_e32 v6, s56
|
|
; VI-NEXT: v_mov_b32_e32 v7, s57
|
|
; VI-NEXT: v_mov_b32_e32 v8, s58
|
|
; VI-NEXT: v_mov_b32_e32 v9, s59
|
|
; VI-NEXT: v_mov_b32_e32 v10, s60
|
|
; VI-NEXT: v_mov_b32_e32 v11, s61
|
|
; VI-NEXT: v_mov_b32_e32 v12, s62
|
|
; VI-NEXT: v_mov_b32_e32 v13, s63
|
|
; VI-NEXT: s_cbranch_execnz .LBB31_3
|
|
; VI-NEXT: .LBB31_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v1, 16
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v49
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v48
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s43, 16
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: v_add_u32_e32 v15, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v39
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_and_b32 s5, s17, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s16, s42, 16
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s5, s16, s5
|
|
; VI-NEXT: s_and_b32 s16, s18, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s17, s41, 16
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v38
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s16, s17, s16
|
|
; VI-NEXT: s_and_b32 s17, s19, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s18, s40, 16
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s17, s18, s17
|
|
; VI-NEXT: s_and_b32 s18, s20, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s15, s15, 16
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: v_add_u32_e32 v17, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v37
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s15, s15, s18
|
|
; VI-NEXT: s_and_b32 s18, s21, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s14, s14, 16
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s14, s14, s18
|
|
; VI-NEXT: s_and_b32 s18, s22, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s13, s13, 16
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v36
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s13, s13, s18
|
|
; VI-NEXT: s_and_b32 s18, s23, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s12, s12, 16
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s12, s12, s18
|
|
; VI-NEXT: s_and_b32 s18, s24, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s11, s11, 16
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: v_add_u32_e32 v19, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v35
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s11, s11, s18
|
|
; VI-NEXT: s_and_b32 s18, s25, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s10, s10, 16
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s10, s10, s18
|
|
; VI-NEXT: s_and_b32 s18, s26, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s9, s9, 16
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v34
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s9, s9, s18
|
|
; VI-NEXT: s_and_b32 s18, s27, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s8, s8, 16
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s8, s8, s18
|
|
; VI-NEXT: s_and_b32 s18, s28, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s7, s7, 16
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: v_add_u32_e32 v21, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v33
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s7, s7, s18
|
|
; VI-NEXT: s_and_b32 s18, s29, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s6, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s6, s6, s18
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v32
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v1, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_add_i32 s4, s4, 0x30000
|
|
; VI-NEXT: s_add_i32 s5, s5, 0x30000
|
|
; VI-NEXT: s_add_i32 s16, s16, 0x30000
|
|
; VI-NEXT: s_add_i32 s17, s17, 0x30000
|
|
; VI-NEXT: s_add_i32 s15, s15, 0x30000
|
|
; VI-NEXT: s_add_i32 s14, s14, 0x30000
|
|
; VI-NEXT: s_add_i32 s13, s13, 0x30000
|
|
; VI-NEXT: s_add_i32 s12, s12, 0x30000
|
|
; VI-NEXT: s_add_i32 s11, s11, 0x30000
|
|
; VI-NEXT: s_add_i32 s10, s10, 0x30000
|
|
; VI-NEXT: s_add_i32 s9, s9, 0x30000
|
|
; VI-NEXT: s_add_i32 s8, s8, 0x30000
|
|
; VI-NEXT: s_add_i32 s7, s7, 0x30000
|
|
; VI-NEXT: s_add_i32 s6, s6, 0x30000
|
|
; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: v_add_u32_e32 v23, vcc, 0x30000, v0
|
|
; 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, s15
|
|
; VI-NEXT: v_mov_b32_e32 v5, s14
|
|
; VI-NEXT: v_mov_b32_e32 v6, s13
|
|
; VI-NEXT: v_mov_b32_e32 v7, s12
|
|
; VI-NEXT: v_mov_b32_e32 v8, s11
|
|
; VI-NEXT: v_mov_b32_e32 v9, s10
|
|
; VI-NEXT: v_mov_b32_e32 v10, s9
|
|
; VI-NEXT: v_mov_b32_e32 v11, s8
|
|
; VI-NEXT: v_mov_b32_e32 v12, s7
|
|
; VI-NEXT: v_mov_b32_e32 v13, s6
|
|
; VI-NEXT: .LBB31_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB31_4:
|
|
; 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_branch .LBB31_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v48i16_to_v24f32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v0
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, 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: buffer_store_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v32
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v33
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v49
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s17, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s18, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s19, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s20, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s21, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s22, s12
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s23, s13
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s24, s14
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s25, s15
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s29, s40
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB31_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v33
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s19
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB31_3
|
|
; GFX9-NEXT: .LBB31_2: ; %cmp.true
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v49
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v48
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v39
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v38
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v37
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v36
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v35
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v34
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v33
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v23
|
|
; GFX9-NEXT: v_pk_add_u16 v0, s6, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, s7, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, s8, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, s9, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, s10, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, s11, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, s12, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, s13, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, s14, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v9, s15, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, s16, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, s17, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, s18, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, s19, 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: .LBB31_3: ; %end
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB31_4:
|
|
; 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_branch .LBB31_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v24f32_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB31_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; 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, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB31_3
|
|
; GFX11-TRUE16-NEXT: .LBB31_2: ; %cmp.true
|
|
; 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, s4, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s17, 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: .LBB31_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB31_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB31_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v48i16_to_v24f32_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s4, s0, s4
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s44
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s16, s5
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s17, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s18, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s19, s8
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s20, s9
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s21, s10
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s22, s11
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s23, s12
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s24, s13
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s25, s14
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s26, s43
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB31_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB31_3
|
|
; GFX11-FAKE16-NEXT: .LBB31_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, s4, 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 v2, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, s17, 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: .LBB31_3: ; %end
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB31_4:
|
|
; GFX11-FAKE16-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
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB31_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <48 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <48 x i16> %a1 to <24 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x i16> %a to <24 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x float> %phi
|
|
}
|
|
|
|
define <48 x half> @bitcast_v24f32_to_v48f16(<24 x float> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v24f32_to_v48f16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v25
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; 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: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; 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: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; 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_lshrrev_b32_e32 v25, 16, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v1
|
|
; SI-NEXT: ; implicit-def: $vgpr1
|
|
; SI-NEXT: ; implicit-def: $vgpr2
|
|
; SI-NEXT: ; implicit-def: $vgpr3
|
|
; SI-NEXT: ; implicit-def: $vgpr4
|
|
; SI-NEXT: ; implicit-def: $vgpr5
|
|
; SI-NEXT: ; implicit-def: $vgpr6
|
|
; 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: $vgpr13
|
|
; SI-NEXT: ; implicit-def: $vgpr14
|
|
; SI-NEXT: ; implicit-def: $vgpr15
|
|
; SI-NEXT: ; implicit-def: $vgpr16
|
|
; SI-NEXT: ; implicit-def: $vgpr17
|
|
; SI-NEXT: ; implicit-def: $vgpr18
|
|
; SI-NEXT: ; implicit-def: $vgpr19
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr22
|
|
; SI-NEXT: ; implicit-def: $vgpr23
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v2, 1.0, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; SI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; SI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; SI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; SI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; SI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; SI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; SI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; SI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; SI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; SI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; SI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; SI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; SI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; SI-NEXT: v_add_f32_e32 v24, 1.0, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v1
|
|
; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v3
|
|
; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v5
|
|
; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v7
|
|
; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v9
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v11
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v13
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v15
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v51
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v49
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v36
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v48
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v52
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v40
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v45
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v47
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v61
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v33
|
|
; SI-NEXT: v_mov_b32_e32 v35, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: .LBB32_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v30
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v3
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
|
|
; SI-NEXT: v_or_b32_e32 v2, v4, v2
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v29
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v28
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v26
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v63
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v25
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v61
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v62
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v59
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v60
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v57
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v58
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v56
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v46
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v43
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v44
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v42
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v41
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v55
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v50
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v53
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v48
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v38
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v36
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v34
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v37
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v39
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v49
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v51
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; 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_v24f32_to_v48f16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; VI-NEXT: .LBB32_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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_lshlrev_b32_e32 v27, 16, v27
|
|
; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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_or_b32_sdwa v20, v20, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v24f32_to_v48f16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v54, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v53, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v52, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v51, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v50, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v49, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v48, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v39, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v37, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v36, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v35, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v34, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v33, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v32, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v31, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v30, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v29, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v28, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v27, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v26, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v25, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s4
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v24f32_to_v48f16:
|
|
; 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, v24
|
|
; 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 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_v24f32_to_v48f16:
|
|
; 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, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v53, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v52, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v51, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v50, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v49, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v48, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v39, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v37, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v35, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v33, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v32, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v31, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v30, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v29, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v28, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v27, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v26, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v25, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v24, v23, 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 <24 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <24 x float> %a1 to <48 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x float> %a to <48 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x half> %phi
|
|
}
|
|
|
|
define inreg <48 x half> @bitcast_v24f32_to_v48f16_scalar(<24 x float> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v24f32_to_v48f16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
|
|
; 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 v18, s16
|
|
; SI-NEXT: v_mov_b32_e32 v15, s17
|
|
; SI-NEXT: v_mov_b32_e32 v13, s18
|
|
; SI-NEXT: v_mov_b32_e32 v12, s19
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_mov_b32_e32 v62, s20
|
|
; SI-NEXT: v_mov_b32_e32 v59, s21
|
|
; SI-NEXT: v_mov_b32_e32 v58, s22
|
|
; SI-NEXT: v_mov_b32_e32 v57, s23
|
|
; SI-NEXT: v_mov_b32_e32 v16, s24
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: v_mov_b32_e32 v17, s25
|
|
; SI-NEXT: v_mov_b32_e32 v14, s26
|
|
; SI-NEXT: v_mov_b32_e32 v11, s27
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v63, s28
|
|
; SI-NEXT: v_mov_b32_e32 v60, s29
|
|
; SI-NEXT: s_cbranch_scc0 .LBB33_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v1
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v60
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, v63
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v60
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v58
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v62
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v17
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v58
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v62
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v12
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v15
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v5
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v4
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v3
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v18
|
|
; SI-NEXT: s_cbranch_execnz .LBB33_3
|
|
; SI-NEXT: .LBB33_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f32_e32 v7, 1.0, v7
|
|
; SI-NEXT: v_add_f32_e32 v1, 1.0, v1
|
|
; SI-NEXT: v_lshrrev_b32_e32 v21, 16, v7
|
|
; SI-NEXT: v_add_f32_e32 v24, 1.0, v57
|
|
; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v21
|
|
; SI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; SI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; 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 v19, 1.0, v62
|
|
; SI-NEXT: v_add_f32_e32 v20, 1.0, v59
|
|
; SI-NEXT: v_add_f32_e32 v22, 1.0, v58
|
|
; SI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; SI-NEXT: v_add_f32_e32 v14, 1.0, v14
|
|
; SI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; SI-NEXT: v_add_f32_e32 v26, 1.0, v63
|
|
; SI-NEXT: v_add_f32_e32 v28, 1.0, v60
|
|
; SI-NEXT: v_add_f32_e32 v2, 1.0, v2
|
|
; SI-NEXT: v_add_f32_e32 v3, 1.0, v3
|
|
; SI-NEXT: v_add_f32_e32 v4, 1.0, v4
|
|
; SI-NEXT: v_add_f32_e32 v5, 1.0, v5
|
|
; SI-NEXT: v_add_f32_e32 v6, 1.0, v6
|
|
; SI-NEXT: v_add_f32_e32 v8, 1.0, v8
|
|
; SI-NEXT: v_add_f32_e32 v9, 1.0, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v15
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v13
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v11
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v26
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v28
|
|
; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v3
|
|
; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v5
|
|
; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v9
|
|
; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v28
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; 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, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v62
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v60
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v58
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v30
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v32
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v36
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v48
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v52
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v40
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v42
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v45
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v47
|
|
; SI-NEXT: v_mov_b32_e32 v21, v10
|
|
; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: .LBB33_3: ; %end
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v61
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v3
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
|
|
; SI-NEXT: v_or_b32_e32 v2, v4, v2
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v42
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v56
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v46
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v44
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v43
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v50
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v41
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v48
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v55
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v38
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v53
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v37
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v51
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v36
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v49
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v34
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v39
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v35
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v33
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v31
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v29
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v24
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v22
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v20
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v23
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v25
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; 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]
|
|
; SI-NEXT: .LBB33_4:
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; kill: killed $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: ; kill: killed $vgpr20
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; kill: killed $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: ; kill: killed $vgpr20
|
|
; SI-NEXT: ; implicit-def: $vgpr19
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr22
|
|
; SI-NEXT: ; implicit-def: $vgpr23
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; kill: killed $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: ; kill: killed $vgpr20
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; kill: killed $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; kill: killed $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; kill: killed $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: s_branch .LBB33_2
|
|
;
|
|
; VI-LABEL: bitcast_v24f32_to_v48f16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, s16
|
|
; VI-NEXT: v_mov_b32_e32 v22, s17
|
|
; VI-NEXT: v_mov_b32_e32 v21, s18
|
|
; VI-NEXT: v_mov_b32_e32 v19, s19
|
|
; VI-NEXT: v_mov_b32_e32 v17, s20
|
|
; VI-NEXT: v_mov_b32_e32 v15, s21
|
|
; VI-NEXT: v_mov_b32_e32 v13, s22
|
|
; VI-NEXT: v_mov_b32_e32 v12, s23
|
|
; VI-NEXT: v_mov_b32_e32 v11, s24
|
|
; VI-NEXT: v_mov_b32_e32 v10, s25
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_mov_b32_e32 v20, s26
|
|
; VI-NEXT: v_mov_b32_e32 v18, s27
|
|
; VI-NEXT: v_mov_b32_e32 v16, s28
|
|
; VI-NEXT: v_mov_b32_e32 v14, s29
|
|
; VI-NEXT: s_cbranch_scc0 .LBB33_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; VI-NEXT: s_cbranch_execnz .LBB33_3
|
|
; VI-NEXT: .LBB33_2: ; %cmp.true
|
|
; 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_add_f32_e32 v14, 1.0, v14
|
|
; VI-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; VI-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; VI-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; VI-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; VI-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; VI-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; VI-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; VI-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; VI-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; VI-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; VI-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; VI-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; VI-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; VI-NEXT: .LBB33_3: ; %end
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_or_b32_sdwa v30, v23, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v31
|
|
; VI-NEXT: v_or_b32_sdwa v31, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v22, 16, v32
|
|
; VI-NEXT: v_or_b32_sdwa v32, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v21, 16, v33
|
|
; VI-NEXT: v_or_b32_sdwa v33, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v24, v17, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v17, 16, v25
|
|
; VI-NEXT: v_or_b32_sdwa v25, v15, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v15, 16, v26
|
|
; VI-NEXT: v_or_b32_sdwa v26, v13, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v27
|
|
; VI-NEXT: v_or_b32_sdwa v27, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v52
|
|
; VI-NEXT: v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v14, 16, v51
|
|
; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v14, v0, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v50
|
|
; VI-NEXT: v_or_b32_sdwa v28, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v53
|
|
; VI-NEXT: v_or_b32_sdwa v15, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v49
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v29
|
|
; VI-NEXT: v_or_b32_sdwa v12, v16, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v48
|
|
; VI-NEXT: v_or_b32_sdwa v29, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v54
|
|
; VI-NEXT: v_or_b32_sdwa v17, v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v39
|
|
; VI-NEXT: v_or_b32_sdwa v11, v18, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v38
|
|
; VI-NEXT: v_lshlrev_b32_e32 v10, 16, v55
|
|
; VI-NEXT: v_or_b32_sdwa v19, v5, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v37
|
|
; VI-NEXT: v_or_b32_sdwa v10, v20, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v6, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v36
|
|
; VI-NEXT: v_or_b32_sdwa v21, v7, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v35
|
|
; VI-NEXT: v_or_b32_sdwa v22, v8, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v34
|
|
; VI-NEXT: v_or_b32_sdwa v23, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, v30
|
|
; VI-NEXT: v_mov_b32_e32 v1, v31
|
|
; VI-NEXT: v_mov_b32_e32 v2, v32
|
|
; VI-NEXT: v_mov_b32_e32 v3, v33
|
|
; VI-NEXT: v_mov_b32_e32 v4, v24
|
|
; VI-NEXT: v_mov_b32_e32 v5, v25
|
|
; VI-NEXT: v_mov_b32_e32 v6, v26
|
|
; VI-NEXT: v_mov_b32_e32 v7, v27
|
|
; VI-NEXT: v_mov_b32_e32 v8, v28
|
|
; VI-NEXT: v_mov_b32_e32 v9, v29
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB33_4:
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr24
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; 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: s_branch .LBB33_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v24f32_to_v48f16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s25
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s27
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s29
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB33_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB33_3
|
|
; GFX9-NEXT: .LBB33_2: ; %cmp.true
|
|
; 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_add_f32_e32 v14, 1.0, v14
|
|
; GFX9-NEXT: v_add_f32_e32 v16, 1.0, v16
|
|
; GFX9-NEXT: v_add_f32_e32 v18, 1.0, v18
|
|
; GFX9-NEXT: v_add_f32_e32 v20, 1.0, v20
|
|
; GFX9-NEXT: v_add_f32_e32 v10, 1.0, v10
|
|
; GFX9-NEXT: v_add_f32_e32 v11, 1.0, v11
|
|
; GFX9-NEXT: v_add_f32_e32 v12, 1.0, v12
|
|
; GFX9-NEXT: v_add_f32_e32 v13, 1.0, v13
|
|
; GFX9-NEXT: v_add_f32_e32 v15, 1.0, v15
|
|
; GFX9-NEXT: v_add_f32_e32 v17, 1.0, v17
|
|
; GFX9-NEXT: v_add_f32_e32 v19, 1.0, v19
|
|
; GFX9-NEXT: v_add_f32_e32 v21, 1.0, v21
|
|
; GFX9-NEXT: v_add_f32_e32 v22, 1.0, v22
|
|
; GFX9-NEXT: v_add_f32_e32 v23, 1.0, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; GFX9-NEXT: .LBB33_3: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v26, 16, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v25, 16, v15
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v50, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v2
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v27, 16, v12
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v49, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v24, 16, v17
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v48, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v4
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v28, v28, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v39, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v33, v33, 16, v19
|
|
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v38, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v6
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v29, v29, 16, v10
|
|
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v37, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v7
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v32, v32, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v36, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v30, v30, 16, v23
|
|
; GFX9-NEXT: v_lshl_or_b32 v31, v31, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v35, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v55, 16, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v54, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v53, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v52, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v34, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, v30
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, v31
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, v32
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, v25
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, v26
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, v27
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, v28
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, v29
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB33_4:
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; 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: s_branch .LBB33_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v24f32_to_v48f16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB33_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB33_3
|
|
; GFX11-TRUE16-NEXT: .LBB33_2: ; %cmp.true
|
|
; 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: v_lshrrev_b32_e32 v24, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; GFX11-TRUE16-NEXT: .LBB33_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v55, v55 :: v_dual_mov_b32 v54, v54
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v53 :: v_dual_mov_b32 v52, v52
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v51, v51 :: v_dual_mov_b32 v50, v50
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v49, v49 :: v_dual_mov_b32 v48, v48
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, v39 :: v_dual_mov_b32 v38, v38
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, v37 :: v_dual_mov_b32 v36, v36
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, v35 :: v_dual_mov_b32 v34, v34
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, v33 :: v_dual_mov_b32 v32, v32
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, v31 :: v_dual_mov_b32 v30, v30
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v29 :: v_dual_mov_b32 v28, v28
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v27 :: v_dual_mov_b32 v26, v26
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v25 :: v_dual_mov_b32 v24, v24
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v31.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v28.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v27.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v26.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v25.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v24.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB33_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB33_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v24f32_to_v48f16_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v23, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v21, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v19, s17
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v7, s20
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s19 :: v_dual_mov_b32 v13, s21
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v9, s25
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v14, s27
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB33_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v6, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB33_3
|
|
; GFX11-FAKE16-NEXT: .LBB33_2: ; %cmp.true
|
|
; 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_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v8, 1.0, v8
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v18, 1.0, v18
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
|
|
; GFX11-FAKE16-NEXT: v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v6, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: .LBB33_3: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v6, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v51, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v49, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v29, v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v39, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v38, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v52, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v48, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v16
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v28, v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v33, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v54, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v53, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v37, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v36, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v34, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v32, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v31, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v30, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v0, v24
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v28
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB33_4:
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr6
|
|
; 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: s_branch .LBB33_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <24 x float> %a, splat (float 1.000000e+00)
|
|
%a2 = bitcast <24 x float> %a1 to <48 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <24 x float> %a to <48 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x half> %phi
|
|
}
|
|
|
|
define <24 x float> @bitcast_v48f16_to_v24f32(<48 x half> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v48f16_to_v24f32:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v14
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:24
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:20
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:32
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:28
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:48
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:44
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:56
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:52
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:64
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:60
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v1
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, v5
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v6
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v11
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, v12
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v15
|
|
; 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_cvt_f16_f32_e32 v0, v20
|
|
; 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_cvt_f16_f32_e32 v0, v23
|
|
; 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_cvt_f16_f32_e32 v0, v22
|
|
; 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_cvt_f16_f32_e32 v0, v25
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v24
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v27
|
|
; 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_cvt_f16_f32_e32 v0, v26
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v29
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v28
|
|
; 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_cvt_f16_f32_e32 v0, v30
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v31
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v40, v63
|
|
; 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_cvt_f16_f32_e32 v0, v41
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v62
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v42
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, v61
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v43
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v60
|
|
; 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_cvt_f16_f32_e32 v0, v44
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v59
|
|
; 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_cvt_f16_f32_e32 v0, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v45, v58
|
|
; 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_cvt_f16_f32_e32 v0, v46
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, v57
|
|
; 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_cvt_f16_f32_e32 v0, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v56
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 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 v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v33
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v55
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v49
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v45
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v43
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v41
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_or_b32_e32 v0, v54, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v52, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v50, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v48, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v38, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v36, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v34, v6
|
|
; SI-NEXT: v_or_b32_e32 v21, v44, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v42, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v40, v23
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; 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: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; 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: buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v20, v46, v20
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; 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: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v55
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v53
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v52
|
|
; 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, v51
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v48
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v38
|
|
; 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, v49
|
|
; 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, v36
|
|
; 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, v34
|
|
; 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, v39
|
|
; 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 v21, v47
|
|
; 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 v22, v46
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; 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, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v44
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v41
|
|
; 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, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v40
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; 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, v33
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v12
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v11, v13, v11
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; 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: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; 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, v14, v12
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; 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: 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_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v15
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_or_b32_e32 v14, v16, v14
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v18
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_or_b32_e32 v17, v19, v17
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v45
|
|
; SI-NEXT: v_or_b32_e32 v20, v22, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v43
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v23, v25, v23
|
|
; 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 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v48f16_to_v24f32:
|
|
; 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: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_mov_b32_e32 v32, v23
|
|
; VI-NEXT: v_mov_b32_e32 v33, v22
|
|
; VI-NEXT: v_mov_b32_e32 v34, v21
|
|
; VI-NEXT: v_mov_b32_e32 v35, v20
|
|
; VI-NEXT: v_mov_b32_e32 v36, v19
|
|
; VI-NEXT: v_mov_b32_e32 v37, v18
|
|
; VI-NEXT: v_mov_b32_e32 v38, v17
|
|
; VI-NEXT: v_mov_b32_e32 v39, v16
|
|
; VI-NEXT: v_mov_b32_e32 v48, v15
|
|
; VI-NEXT: v_mov_b32_e32 v49, v14
|
|
; VI-NEXT: v_mov_b32_e32 v50, v13
|
|
; VI-NEXT: v_mov_b32_e32 v51, v12
|
|
; VI-NEXT: v_mov_b32_e32 v52, v11
|
|
; VI-NEXT: v_mov_b32_e32 v53, v10
|
|
; VI-NEXT: v_mov_b32_e32 v54, v9
|
|
; VI-NEXT: v_mov_b32_e32 v55, v8
|
|
; VI-NEXT: v_mov_b32_e32 v40, v7
|
|
; VI-NEXT: v_mov_b32_e32 v41, v6
|
|
; VI-NEXT: v_mov_b32_e32 v42, v5
|
|
; VI-NEXT: v_mov_b32_e32 v43, v4
|
|
; VI-NEXT: v_mov_b32_e32 v44, v3
|
|
; VI-NEXT: v_mov_b32_e32 v45, v2
|
|
; VI-NEXT: v_mov_b32_e32 v46, v1
|
|
; VI-NEXT: v_mov_b32_e32 v47, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 v23, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v23, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v23, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v23, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v23, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v23, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v23, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v23, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v23, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v23, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v23, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v23, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v23, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v23, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v23, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v23, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v23, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v23, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v23, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v23, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v23, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v23, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v23, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v23, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v44, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v43, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v42, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v41, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v40, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v55, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v53, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v52, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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 v23, 0x200
|
|
; VI-NEXT: v_add_f16_sdwa v0, v47, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v47
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_add_f16_sdwa v1, v46, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v2, 0x200, v46
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; VI-NEXT: v_add_f16_sdwa v2, v45, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v45
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v3, v44, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v44
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_add_f16_sdwa v4, v43, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v43
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_add_f16_sdwa v5, v42, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v42
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_add_f16_sdwa v6, v41, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v41
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_add_f16_sdwa v7, v40, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v40
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_add_f16_sdwa v8, v55, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v55
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_add_f16_sdwa v9, v54, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, 0x200, v54
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, 0x200, v53
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_add_f16_sdwa v11, v52, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, 0x200, v52
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_add_f16_sdwa v12, v51, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, 0x200, v51
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_add_f16_sdwa v13, v50, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v50
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_add_f16_sdwa v14, v49, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v48, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v39, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v38, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v37, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v36, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v35, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v34, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB34_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48f16_to_v24f32:
|
|
; 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 v33, v22
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v13
|
|
; 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, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v12
|
|
; 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, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v11
|
|
; 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, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v10
|
|
; 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, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v1
|
|
; 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, v53
|
|
; 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, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v47
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 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:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; 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: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; 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 v9, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: s_movk_i32 s7, 0x200
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, 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(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, 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: .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_v48f16_to_v24f32:
|
|
; 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, v24
|
|
; 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: .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_v48f16_to_v24f32:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v24
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v53, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v55, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v64, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v52, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v51, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v50, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v49, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v48, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v39, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v38, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v37, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v36, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v35, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v34, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v33, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v32, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v31, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v30, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v29, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v28, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v27, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v26, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v25, v23, 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: .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 <48 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <48 x half> %a1 to <24 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x half> %a to <24 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x float> %phi
|
|
}
|
|
|
|
define inreg <24 x float> @bitcast_v48f16_to_v24f32_scalar(<48 x half> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v48f16_to_v24f32_scalar:
|
|
; 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:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v62, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v63, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v61, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v60, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v59, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v58, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v57, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v56, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v45, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, s17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, s16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, s19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, s18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, s21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, s20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, s23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, s22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, s25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, s24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, s27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v30, s28
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v31
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v34
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v35
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, s26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, s29
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v40
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: s_cbranch_scc0 .LBB35_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v0, v8, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v7, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v6, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v5, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v34
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v62
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v52
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v48
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v60
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v58
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v46
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v4, v54, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v35, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v30, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v55, v7
|
|
; SI-NEXT: v_or_b32_e32 v8, v63, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v43, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v49, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v39, v11
|
|
; SI-NEXT: v_or_b32_e32 v12, v37, v12
|
|
; SI-NEXT: v_or_b32_e32 v13, v61, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v33, v14
|
|
; SI-NEXT: v_or_b32_e32 v15, v59, v15
|
|
; SI-NEXT: v_or_b32_e32 v16, v57, v16
|
|
; SI-NEXT: v_or_b32_e32 v17, v47, v17
|
|
; SI-NEXT: v_or_b32_e32 v18, v45, v18
|
|
; SI-NEXT: v_or_b32_e32 v19, v25, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v41, v20
|
|
; SI-NEXT: v_or_b32_e32 v21, v28, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v26, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v31, v23
|
|
; SI-NEXT: s_cbranch_execnz .LBB35_3
|
|
; SI-NEXT: .LBB35_2: ; %cmp.true
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v30
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v55
|
|
; 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_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; 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 v9, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v49
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v48
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v60
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v33
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; 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_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v59
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v56
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v47
|
|
; 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, v45
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v42
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v41
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v28
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, 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 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: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_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 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; 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, v51
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; 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, v34
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; 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, v62
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; 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, v53
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; 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, v52
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; 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, v50
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v38
|
|
; SI-NEXT: v_or_b32_e32 v11, v13, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v36
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v61
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v32
|
|
; SI-NEXT: v_or_b32_e32 v14, v16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v58
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v57
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v46
|
|
; SI-NEXT: v_or_b32_e32 v17, v19, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v44
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v31
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v29
|
|
; SI-NEXT: v_or_b32_e32 v20, v22, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v27
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v23, v25, v23
|
|
; SI-NEXT: .LBB35_3: ; %end
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB35_4:
|
|
; SI-NEXT: v_mov_b32_e32 v40, v31
|
|
; SI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v54, v33
|
|
; SI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v51, v32
|
|
; SI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v34, v59
|
|
; SI-NEXT: v_mov_b32_e32 v59, v46
|
|
; SI-NEXT: v_mov_b32_e32 v46, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v24
|
|
; SI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v35, v60
|
|
; SI-NEXT: v_mov_b32_e32 v60, v47
|
|
; SI-NEXT: v_mov_b32_e32 v47, v42
|
|
; SI-NEXT: v_mov_b32_e32 v42, v26
|
|
; SI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v55, v52
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_mov_b32_e32 v53, v43
|
|
; SI-NEXT: v_mov_b32_e32 v52, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v61
|
|
; SI-NEXT: v_mov_b32_e32 v61, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v25
|
|
; SI-NEXT: v_mov_b32_e32 v43, v27
|
|
; SI-NEXT: v_mov_b32_e32 v33, v62
|
|
; SI-NEXT: v_mov_b32_e32 v62, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v44
|
|
; SI-NEXT: v_mov_b32_e32 v44, v28
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v63, v58
|
|
; SI-NEXT: v_mov_b32_e32 v58, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v29
|
|
; SI-NEXT: v_mov_b32_e32 v32, v30
|
|
; 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: v_mov_b32_e32 v29, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v58
|
|
; SI-NEXT: v_mov_b32_e32 v58, v63
|
|
; SI-NEXT: v_mov_b32_e32 v27, v43
|
|
; SI-NEXT: v_mov_b32_e32 v25, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v61
|
|
; SI-NEXT: v_mov_b32_e32 v61, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v52
|
|
; SI-NEXT: v_mov_b32_e32 v43, v53
|
|
; SI-NEXT: v_mov_b32_e32 v52, v55
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v30, v32
|
|
; SI-NEXT: v_mov_b32_e32 v28, v44
|
|
; SI-NEXT: v_mov_b32_e32 v44, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v62
|
|
; SI-NEXT: v_mov_b32_e32 v62, v33
|
|
; SI-NEXT: v_mov_b32_e32 v26, v42
|
|
; SI-NEXT: v_mov_b32_e32 v42, v47
|
|
; SI-NEXT: v_mov_b32_e32 v47, v60
|
|
; SI-NEXT: v_mov_b32_e32 v60, v35
|
|
; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v24, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v46
|
|
; SI-NEXT: v_mov_b32_e32 v46, v59
|
|
; SI-NEXT: v_mov_b32_e32 v59, v34
|
|
; SI-NEXT: v_mov_b32_e32 v32, v51
|
|
; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v33, v54
|
|
; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v31, v40
|
|
; SI-NEXT: s_branch .LBB35_2
|
|
;
|
|
; VI-LABEL: bitcast_v48f16_to_v24f32_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v32, v9
|
|
; VI-NEXT: v_mov_b32_e32 v33, v8
|
|
; VI-NEXT: v_mov_b32_e32 v34, v7
|
|
; VI-NEXT: v_mov_b32_e32 v35, v6
|
|
; VI-NEXT: v_mov_b32_e32 v36, v5
|
|
; VI-NEXT: v_mov_b32_e32 v37, v4
|
|
; VI-NEXT: v_mov_b32_e32 v38, v3
|
|
; VI-NEXT: v_mov_b32_e32 v39, v2
|
|
; VI-NEXT: v_mov_b32_e32 v48, v1
|
|
; VI-NEXT: v_mov_b32_e32 v49, v0
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: s_cbranch_scc0 .LBB35_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s43, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s44, s42, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s44
|
|
; VI-NEXT: s_and_b32 s44, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s45, s41, 16
|
|
; VI-NEXT: s_or_b32 s44, s44, s45
|
|
; VI-NEXT: s_and_b32 s45, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s46, s40, 16
|
|
; VI-NEXT: s_or_b32 s45, s45, s46
|
|
; VI-NEXT: s_and_b32 s46, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s47, s15, 16
|
|
; VI-NEXT: s_or_b32 s46, s46, s47
|
|
; VI-NEXT: s_and_b32 s47, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s56, s14, 16
|
|
; VI-NEXT: v_mov_b32_e32 v0, 16
|
|
; VI-NEXT: s_or_b32 s47, s47, s56
|
|
; VI-NEXT: s_and_b32 s56, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s57, s13, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s56, s56, s57
|
|
; VI-NEXT: s_and_b32 s57, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s58, s12, 16
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s57, s57, s58
|
|
; VI-NEXT: s_and_b32 s58, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s59, s11, 16
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s58, s58, s59
|
|
; VI-NEXT: s_and_b32 s59, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s60, s10, 16
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s59, s59, s60
|
|
; VI-NEXT: s_and_b32 s60, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s61, s9, 16
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s60, s60, s61
|
|
; VI-NEXT: s_and_b32 s61, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s62, s8, 16
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s61, s61, s62
|
|
; VI-NEXT: s_and_b32 s62, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s63, s7, 16
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s62, s62, s63
|
|
; VI-NEXT: s_and_b32 s63, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s72, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s63, s63, s72
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v0, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s44
|
|
; VI-NEXT: v_mov_b32_e32 v3, s45
|
|
; VI-NEXT: v_mov_b32_e32 v4, s46
|
|
; VI-NEXT: v_mov_b32_e32 v5, s47
|
|
; VI-NEXT: v_mov_b32_e32 v6, s56
|
|
; VI-NEXT: v_mov_b32_e32 v7, s57
|
|
; VI-NEXT: v_mov_b32_e32 v8, s58
|
|
; VI-NEXT: v_mov_b32_e32 v9, s59
|
|
; VI-NEXT: v_mov_b32_e32 v10, s60
|
|
; VI-NEXT: v_mov_b32_e32 v11, s61
|
|
; VI-NEXT: v_mov_b32_e32 v12, s62
|
|
; VI-NEXT: v_mov_b32_e32 v13, s63
|
|
; VI-NEXT: s_cbranch_execnz .LBB35_3
|
|
; VI-NEXT: .LBB35_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v13, 0x200
|
|
; VI-NEXT: v_mov_b32_e32 v0, s43
|
|
; VI-NEXT: v_add_f16_sdwa v0, v0, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, s16, v13
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_mov_b32_e32 v1, s42
|
|
; VI-NEXT: v_add_f16_sdwa v1, v1, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v2, s17, v13
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; VI-NEXT: v_mov_b32_e32 v2, s41
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s18, v13
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v3, s40
|
|
; VI-NEXT: v_add_f16_sdwa v3, v3, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, s19, v13
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_mov_b32_e32 v4, s15
|
|
; VI-NEXT: v_add_f16_sdwa v4, v4, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, s20, v13
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_mov_b32_e32 v5, s14
|
|
; VI-NEXT: v_add_f16_sdwa v5, v5, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, s21, v13
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_mov_b32_e32 v6, s13
|
|
; VI-NEXT: v_add_f16_sdwa v6, v6, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, s22, v13
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_mov_b32_e32 v7, s12
|
|
; VI-NEXT: v_add_f16_sdwa v7, v7, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, s23, v13
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_mov_b32_e32 v8, s11
|
|
; VI-NEXT: v_add_f16_sdwa v8, v8, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, s24, v13
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_mov_b32_e32 v9, s10
|
|
; VI-NEXT: v_add_f16_sdwa v9, v9, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, s25, v13
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_mov_b32_e32 v10, s9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v10, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, s26, v13
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_mov_b32_e32 v11, s8
|
|
; VI-NEXT: v_add_f16_sdwa v11, v11, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, s27, v13
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_mov_b32_e32 v12, s7
|
|
; VI-NEXT: v_add_f16_sdwa v12, v12, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, s28, v13
|
|
; VI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; VI-NEXT: v_mov_b32_e32 v14, s6
|
|
; VI-NEXT: v_add_f16_sdwa v14, v14, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, s29, v13
|
|
; VI-NEXT: v_mov_b32_e32 v23, 0x200
|
|
; VI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; VI-NEXT: v_add_f16_sdwa v14, v49, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v48, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v39, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v38, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v37, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v36, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v35, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v34, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB35_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB35_4:
|
|
; 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_branch .LBB35_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v48f16_to_v24f32_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v0
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, 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: buffer_store_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v32
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v33
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v49
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s17, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s18, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s19, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s20, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s21, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s22, s12
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s23, s13
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s24, s14
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s25, s15
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s29, s40
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB35_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v33
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s19
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB35_3
|
|
; GFX9-NEXT: .LBB35_2: ; %cmp.true
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v49
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v48
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v39
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v38
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v37
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v36
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v35
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v34
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v33
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v32
|
|
; GFX9-NEXT: s_movk_i32 s4, 0x200
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, 0x200
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v23
|
|
; GFX9-NEXT: v_pk_add_f16 v0, s6, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, s7, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, s8, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, s9, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, s10, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, s11, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, s12, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, s13, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, s14, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v9, s15, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, s16, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, s17, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, s18, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, s19, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, v14, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, v15, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, v16, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, v17, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, v18, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, v19, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, v20, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, v21, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, v22, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, v23, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB35_3: ; %end
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB35_4:
|
|
; 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_branch .LBB35_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v24f32_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB35_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; 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, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB35_3
|
|
; GFX11-TRUE16-NEXT: .LBB35_2: ; %cmp.true
|
|
; 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, s4 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, s17 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: .LBB35_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB35_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB35_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v48f16_to_v24f32_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s4, s0, s4
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s44
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s16, s5
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s17, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s18, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s19, s8
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s20, s9
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s21, s10
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s22, s11
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s23, s12
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s24, s13
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s25, s14
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s26, s43
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB35_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB35_3
|
|
; GFX11-FAKE16-NEXT: .LBB35_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, s4 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 v2, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, s17 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: .LBB35_3: ; %end
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB35_4:
|
|
; GFX11-FAKE16-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
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB35_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <48 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <48 x half> %a1 to <24 x float>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x half> %a to <24 x float>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <24 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <24 x float> %phi
|
|
}
|
|
|
|
define <12 x double> @bitcast_v12i64_to_v12f64(<12 x i64> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v12i64_to_v12f64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: .LBB36_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v12i64_to_v12f64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: .LBB36_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v12i64_to_v12f64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: .LBB36_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v12i64_to_v12f64:
|
|
; 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, v24
|
|
; 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: .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 <12 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <12 x i64> %a1 to <12 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x i64> %a to <12 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x double> %phi
|
|
}
|
|
|
|
define inreg <12 x double> @bitcast_v12i64_to_v12f64_scalar(<12 x i64> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v12i64_to_v12f64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB37_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB37_4:
|
|
; SI-NEXT: s_branch .LBB37_2
|
|
;
|
|
; VI-LABEL: bitcast_v12i64_to_v12f64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB37_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB37_4:
|
|
; VI-NEXT: s_branch .LBB37_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v12i64_to_v12f64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; 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: 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: 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: .LBB37_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB37_4:
|
|
; GFX9-NEXT: s_branch .LBB37_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v12i64_to_v12f64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB37_3
|
|
; GFX11-NEXT: ; %bb.1: ; %Flow
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-NEXT: s_cbranch_vccz .LBB37_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB37_3:
|
|
; GFX11-NEXT: .LBB37_4: ; %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: 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 <12 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <12 x i64> %a1 to <12 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x i64> %a to <12 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x double> %phi
|
|
}
|
|
|
|
define <12 x i64> @bitcast_v12f64_to_v12i64(<12 x double> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v12f64_to_v12i64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: .LBB38_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v12f64_to_v12i64:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: .LBB38_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v12f64_to_v12i64:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: .LBB38_2: ; %end
|
|
; GFX9-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-LABEL: bitcast_v12f64_to_v12i64:
|
|
; 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, v24
|
|
; 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: .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 <12 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <12 x double> %a1 to <12 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x double> %a to <12 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x i64> %phi
|
|
}
|
|
|
|
define inreg <12 x i64> @bitcast_v12f64_to_v12i64_scalar(<12 x double> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v12f64_to_v12i64_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_mov_b32_e32 v23, v9
|
|
; SI-NEXT: v_mov_b32_e32 v22, v8
|
|
; SI-NEXT: v_mov_b32_e32 v21, v7
|
|
; SI-NEXT: v_mov_b32_e32 v20, v6
|
|
; SI-NEXT: v_mov_b32_e32 v19, v5
|
|
; SI-NEXT: v_mov_b32_e32 v18, v4
|
|
; SI-NEXT: v_mov_b32_e32 v17, v3
|
|
; SI-NEXT: v_mov_b32_e32 v16, v2
|
|
; SI-NEXT: v_mov_b32_e32 v15, v1
|
|
; SI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: v_mov_b32_e32 v12, s28
|
|
; SI-NEXT: v_mov_b32_e32 v13, s29
|
|
; SI-NEXT: s_cbranch_scc0 .LBB39_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB39_3
|
|
; SI-NEXT: .LBB39_2: ; %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: .LBB39_3: ; %end
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB39_4:
|
|
; SI-NEXT: s_branch .LBB39_2
|
|
;
|
|
; VI-LABEL: bitcast_v12f64_to_v12i64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, v9
|
|
; VI-NEXT: v_mov_b32_e32 v22, v8
|
|
; VI-NEXT: v_mov_b32_e32 v21, v7
|
|
; VI-NEXT: v_mov_b32_e32 v20, v6
|
|
; VI-NEXT: v_mov_b32_e32 v19, v5
|
|
; VI-NEXT: v_mov_b32_e32 v18, v4
|
|
; VI-NEXT: v_mov_b32_e32 v17, v3
|
|
; VI-NEXT: v_mov_b32_e32 v16, v2
|
|
; VI-NEXT: v_mov_b32_e32 v15, v1
|
|
; VI-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: s_cbranch_scc0 .LBB39_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_cbranch_execnz .LBB39_3
|
|
; VI-NEXT: .LBB39_2: ; %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: .LBB39_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB39_4:
|
|
; VI-NEXT: s_branch .LBB39_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v12f64_to_v12i64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, v0
|
|
; 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: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s29
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB39_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB39_3
|
|
; GFX9-NEXT: .LBB39_2: ; %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: .LBB39_3: ; %end
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB39_4:
|
|
; GFX9-NEXT: s_branch .LBB39_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v12f64_to_v12i64_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; 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_vccz .LBB39_4
|
|
; GFX11-NEXT: ; %bb.2: ; %end
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB39_3:
|
|
; GFX11-NEXT: .LBB39_4: ; %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: 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 <12 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <12 x double> %a1 to <12 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x double> %a to <12 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x i64> %phi
|
|
}
|
|
|
|
define <48 x i16> @bitcast_v12i64_to_v48i16(<12 x i64> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v12i64_to_v48i16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v25
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; 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 v25, v24, v23, 16
|
|
; SI-NEXT: v_alignbit_b32 v26, v22, v21, 16
|
|
; SI-NEXT: v_alignbit_b32 v27, v20, v19, 16
|
|
; SI-NEXT: v_alignbit_b32 v28, v18, v17, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v16, v15, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v14, v13, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v12, v11, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v10, v9, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v8, v7, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v6, v5, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v4, v3, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v2, v1, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v2
|
|
; 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 v1, vcc, 3, v1
|
|
; SI-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
|
|
; SI-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; SI-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
|
|
; SI-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_addc_u32_e32 v18, vcc, 0, v18, vcc
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_addc_u32_e32 v20, vcc, 0, v20, vcc
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_addc_u32_e32 v22, vcc, 0, v22, vcc
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_addc_u32_e32 v24, vcc, 0, v24, vcc
|
|
; SI-NEXT: v_alignbit_b32 v25, v24, v23, 16
|
|
; SI-NEXT: v_alignbit_b32 v26, v22, v21, 16
|
|
; SI-NEXT: v_alignbit_b32 v27, v20, v19, 16
|
|
; SI-NEXT: v_alignbit_b32 v28, v18, v17, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v16, v15, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v14, v13, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v12, v11, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v10, v9, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v8, v7, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v6, v5, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v4, v3, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v2, v1, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v2
|
|
; SI-NEXT: .LBB40_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v53
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 4, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 8, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 12, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 16, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 20, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 24, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 28, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 32, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 36, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 40, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v31
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 52, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v29
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 56, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 60, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 64, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v27
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x4c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x50, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x54, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x58, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v30
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v12i64_to_v48i16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; VI-NEXT: .LBB40_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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_lshlrev_b32_e32 v27, 16, v27
|
|
; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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_or_b32_sdwa v20, v20, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v12i64_to_v48i16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v54, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v53, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v52, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v51, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v50, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v49, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v48, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v39, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v37, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v36, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v35, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v34, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v33, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v32, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v31, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v30, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v29, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v28, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v27, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v26, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v25, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s4
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v12i64_to_v48i16:
|
|
; 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, v24
|
|
; 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 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_v12i64_to_v48i16:
|
|
; 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, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v53, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v52, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v51, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v50, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v49, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v48, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v39, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v37, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v35, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v33, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v32, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v31, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v30, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v29, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v28, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v27, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v26, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v25, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v24, v23, 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 <12 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <12 x i64> %a1 to <48 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x i64> %a to <48 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x i16> %phi
|
|
}
|
|
|
|
define inreg <48 x i16> @bitcast_v12i64_to_v48i16_scalar(<12 x i64> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v12i64_to_v48i16_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 v20, 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 v20, s30, 0
|
|
; SI-NEXT: v_mov_b32_e32 v12, s16
|
|
; SI-NEXT: v_mov_b32_e32 v13, s17
|
|
; SI-NEXT: v_mov_b32_e32 v14, s18
|
|
; SI-NEXT: v_mov_b32_e32 v15, s19
|
|
; SI-NEXT: v_mov_b32_e32 v16, s20
|
|
; SI-NEXT: v_mov_b32_e32 v17, s21
|
|
; SI-NEXT: v_writelane_b32 v20, s31, 1
|
|
; SI-NEXT: v_mov_b32_e32 v18, s22
|
|
; SI-NEXT: v_mov_b32_e32 v19, s23
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v12
|
|
; SI-NEXT: v_mov_b32_e32 v12, s24
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v13
|
|
; SI-NEXT: v_mov_b32_e32 v13, s25
|
|
; SI-NEXT: v_readfirstlane_b32 s24, v14
|
|
; SI-NEXT: v_mov_b32_e32 v14, s26
|
|
; SI-NEXT: v_readfirstlane_b32 s25, v15
|
|
; SI-NEXT: v_mov_b32_e32 v15, s27
|
|
; SI-NEXT: v_readfirstlane_b32 s22, v16
|
|
; SI-NEXT: v_mov_b32_e32 v16, s28
|
|
; SI-NEXT: v_readfirstlane_b32 s23, v17
|
|
; SI-NEXT: v_mov_b32_e32 v17, s29
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
|
|
; SI-NEXT: v_writelane_b32 v20, s34, 2
|
|
; SI-NEXT: v_readfirstlane_b32 s20, v18
|
|
; SI-NEXT: v_readfirstlane_b32 s21, v19
|
|
; SI-NEXT: v_readfirstlane_b32 s18, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s19, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s16, v14
|
|
; SI-NEXT: v_readfirstlane_b32 s17, v15
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v16
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v17
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s4, v9
|
|
; SI-NEXT: s_and_b64 s[26:27], vcc, exec
|
|
; SI-NEXT: v_readfirstlane_b32 s5, v10
|
|
; SI-NEXT: v_writelane_b32 v20, s35, 3
|
|
; SI-NEXT: s_cbranch_scc0 .LBB41_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s88, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s90, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s91, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s92, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s94, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s95, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s41, 16
|
|
; SI-NEXT: s_lshr_b64 s[26:27], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[28:29], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[16:17], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[40:41], 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 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 s40, s40, 3
|
|
; SI-NEXT: s_addc_u32 s41, s41, 0
|
|
; SI-NEXT: s_lshr_b32 s88, s5, 16
|
|
; SI-NEXT: s_lshr_b32 s89, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s90, s9, 16
|
|
; SI-NEXT: s_lshr_b32 s91, s11, 16
|
|
; SI-NEXT: s_lshr_b32 s92, s13, 16
|
|
; SI-NEXT: s_lshr_b32 s93, s15, 16
|
|
; SI-NEXT: s_lshr_b32 s94, s17, 16
|
|
; SI-NEXT: s_lshr_b32 s95, s19, 16
|
|
; SI-NEXT: s_lshr_b32 s30, s21, 16
|
|
; SI-NEXT: s_lshr_b32 s31, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s34, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s35, s41, 16
|
|
; SI-NEXT: s_lshr_b64 s[26:27], s[4:5], 16
|
|
; SI-NEXT: s_lshr_b64 s[28:29], s[6:7], 16
|
|
; SI-NEXT: s_lshr_b64 s[42:43], s[8:9], 16
|
|
; SI-NEXT: s_lshr_b64 s[44:45], s[10:11], 16
|
|
; SI-NEXT: s_lshr_b64 s[46:47], s[12:13], 16
|
|
; SI-NEXT: s_lshr_b64 s[56:57], s[14:15], 16
|
|
; SI-NEXT: s_lshr_b64 s[58:59], s[16:17], 16
|
|
; SI-NEXT: s_lshr_b64 s[60:61], s[18:19], 16
|
|
; SI-NEXT: s_lshr_b64 s[62:63], s[20:21], 16
|
|
; SI-NEXT: s_lshr_b64 s[72:73], s[22:23], 16
|
|
; SI-NEXT: s_lshr_b64 s[74:75], s[24:25], 16
|
|
; SI-NEXT: s_lshr_b64 s[76:77], s[40:41], 16
|
|
; SI-NEXT: .LBB41_3: ; %end
|
|
; SI-NEXT: s_lshl_b32 s27, s76, 16
|
|
; SI-NEXT: s_and_b32 s29, s40, 0xffff
|
|
; SI-NEXT: s_or_b32 s27, s29, s27
|
|
; SI-NEXT: v_mov_b32_e32 v1, s27
|
|
; SI-NEXT: s_and_b32 s27, s41, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s29, s35, 16
|
|
; SI-NEXT: s_or_b32 s27, s27, s29
|
|
; SI-NEXT: v_mov_b32_e32 v2, s27
|
|
; SI-NEXT: s_lshl_b32 s27, s74, 16
|
|
; SI-NEXT: s_and_b32 s24, s24, 0xffff
|
|
; SI-NEXT: s_or_b32 s24, s24, s27
|
|
; SI-NEXT: v_mov_b32_e32 v3, s24
|
|
; SI-NEXT: s_and_b32 s24, s25, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s25, s34, 16
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
|
|
; SI-NEXT: s_or_b32 s24, s24, s25
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 8, v0
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s24
|
|
; SI-NEXT: s_and_b32 s22, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s24, s72, 16
|
|
; SI-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 12, v0
|
|
; SI-NEXT: s_or_b32 s22, s22, s24
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s22
|
|
; SI-NEXT: s_and_b32 s22, s23, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s23, s31, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 16, v0
|
|
; SI-NEXT: s_or_b32 s22, s22, s23
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s22
|
|
; SI-NEXT: s_and_b32 s20, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s22, s62, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 20, v0
|
|
; SI-NEXT: s_or_b32 s20, s20, s22
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s20
|
|
; SI-NEXT: s_and_b32 s20, s21, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s21, s30, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 24, v0
|
|
; SI-NEXT: s_or_b32 s20, s20, s21
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s20
|
|
; SI-NEXT: s_and_b32 s18, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s20, s60, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 28, v0
|
|
; SI-NEXT: s_or_b32 s18, s18, s20
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: s_and_b32 s18, s19, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s19, s95, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 32, v0
|
|
; SI-NEXT: s_or_b32 s18, s18, s19
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s18
|
|
; SI-NEXT: s_and_b32 s16, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s18, s58, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 36, v0
|
|
; SI-NEXT: s_or_b32 s16, s16, s18
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s16
|
|
; SI-NEXT: s_and_b32 s16, s17, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s17, s94, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 40, v0
|
|
; SI-NEXT: s_or_b32 s16, s16, s17
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s16
|
|
; SI-NEXT: s_and_b32 s14, s14, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s16, s56, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 44, v0
|
|
; SI-NEXT: s_or_b32 s14, s14, s16
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s14
|
|
; SI-NEXT: s_and_b32 s14, s15, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s15, s93, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 48, v0
|
|
; SI-NEXT: s_or_b32 s14, s14, s15
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s14
|
|
; SI-NEXT: s_and_b32 s12, s12, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s14, s46, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 52, v0
|
|
; SI-NEXT: s_or_b32 s12, s12, s14
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s12
|
|
; SI-NEXT: s_and_b32 s12, s13, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s13, s92, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 56, v0
|
|
; SI-NEXT: s_or_b32 s12, s12, s13
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s12
|
|
; SI-NEXT: s_and_b32 s10, s10, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s12, s44, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 60, v0
|
|
; SI-NEXT: s_or_b32 s10, s10, s12
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s10
|
|
; SI-NEXT: s_and_b32 s10, s11, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s11, s91, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 64, v0
|
|
; SI-NEXT: s_or_b32 s10, s10, s11
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s10
|
|
; SI-NEXT: s_and_b32 s8, s8, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s10, s42, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x44, v0
|
|
; SI-NEXT: s_or_b32 s8, s8, s10
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s8
|
|
; SI-NEXT: s_and_b32 s8, s9, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s90, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x48, v0
|
|
; SI-NEXT: s_or_b32 s8, s8, s9
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s8
|
|
; SI-NEXT: s_and_b32 s6, s6, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s28, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x4c, v0
|
|
; SI-NEXT: s_or_b32 s6, s6, s8
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: s_and_b32 s6, s7, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s89, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x50, v0
|
|
; SI-NEXT: s_or_b32 s6, s6, s7
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: s_and_b32 s4, s4, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s26, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x54, v0
|
|
; SI-NEXT: s_or_b32 s4, s4, s6
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v2, s4
|
|
; SI-NEXT: s_and_b32 s4, s5, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s88, 16
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x58, v0
|
|
; SI-NEXT: s_or_b32 s4, s4, s5
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_mov_b32_e32 v1, s4
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: v_readlane_b32 s35, v20, 3
|
|
; SI-NEXT: v_readlane_b32 s34, v20, 2
|
|
; SI-NEXT: v_readlane_b32 s31, v20, 1
|
|
; SI-NEXT: v_readlane_b32 s30, v20, 0
|
|
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_mov_b64 exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB41_4:
|
|
; SI-NEXT: ; implicit-def: $sgpr76
|
|
; SI-NEXT: ; implicit-def: $sgpr35
|
|
; SI-NEXT: ; implicit-def: $sgpr74
|
|
; SI-NEXT: ; implicit-def: $sgpr34
|
|
; SI-NEXT: ; implicit-def: $sgpr72
|
|
; SI-NEXT: ; implicit-def: $sgpr31
|
|
; SI-NEXT: ; implicit-def: $sgpr62
|
|
; SI-NEXT: ; implicit-def: $sgpr30
|
|
; SI-NEXT: ; implicit-def: $sgpr60
|
|
; SI-NEXT: ; implicit-def: $sgpr95
|
|
; SI-NEXT: ; implicit-def: $sgpr58
|
|
; SI-NEXT: ; implicit-def: $sgpr94
|
|
; SI-NEXT: ; implicit-def: $sgpr56
|
|
; SI-NEXT: ; implicit-def: $sgpr93
|
|
; SI-NEXT: ; implicit-def: $sgpr46
|
|
; SI-NEXT: ; implicit-def: $sgpr92
|
|
; SI-NEXT: ; implicit-def: $sgpr44
|
|
; SI-NEXT: ; implicit-def: $sgpr91
|
|
; SI-NEXT: ; implicit-def: $sgpr42
|
|
; SI-NEXT: ; implicit-def: $sgpr90
|
|
; SI-NEXT: ; implicit-def: $sgpr28
|
|
; SI-NEXT: ; implicit-def: $sgpr89
|
|
; SI-NEXT: ; implicit-def: $sgpr26
|
|
; SI-NEXT: ; implicit-def: $sgpr88
|
|
; SI-NEXT: s_branch .LBB41_2
|
|
;
|
|
; VI-LABEL: bitcast_v12i64_to_v48i16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_mov_b32_e32 v11, s16
|
|
; VI-NEXT: v_mov_b32_e32 v12, s17
|
|
; VI-NEXT: v_mov_b32_e32 v13, s18
|
|
; VI-NEXT: v_mov_b32_e32 v14, s19
|
|
; VI-NEXT: v_mov_b32_e32 v15, s20
|
|
; VI-NEXT: v_mov_b32_e32 v16, s21
|
|
; VI-NEXT: v_mov_b32_e32 v17, s22
|
|
; VI-NEXT: v_mov_b32_e32 v18, s23
|
|
; VI-NEXT: v_mov_b32_e32 v19, s24
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v11
|
|
; VI-NEXT: v_mov_b32_e32 v11, s25
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v12
|
|
; VI-NEXT: v_mov_b32_e32 v12, s26
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v13
|
|
; VI-NEXT: v_mov_b32_e32 v13, s27
|
|
; VI-NEXT: v_readfirstlane_b32 s26, v14
|
|
; VI-NEXT: v_mov_b32_e32 v14, s28
|
|
; VI-NEXT: v_readfirstlane_b32 s25, v15
|
|
; VI-NEXT: v_mov_b32_e32 v15, s29
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s24, v16
|
|
; VI-NEXT: v_readfirstlane_b32 s23, v17
|
|
; VI-NEXT: v_readfirstlane_b32 s22, v18
|
|
; VI-NEXT: v_readfirstlane_b32 s21, v19
|
|
; VI-NEXT: v_readfirstlane_b32 s20, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s19, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s18, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s17, v14
|
|
; VI-NEXT: v_readfirstlane_b32 s16, v15
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v0
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v8
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v9
|
|
; VI-NEXT: s_cbranch_scc0 .LBB41_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s27, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s28, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s29, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s43, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s44, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s42, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB41_3
|
|
; VI-NEXT: .LBB41_2: ; %cmp.true
|
|
; VI-NEXT: s_add_u32 s6, s6, 3
|
|
; VI-NEXT: s_addc_u32 s7, s7, 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 s17, s17, 3
|
|
; VI-NEXT: s_addc_u32 s16, s16, 0
|
|
; VI-NEXT: s_add_u32 s19, s19, 3
|
|
; VI-NEXT: s_addc_u32 s18, s18, 0
|
|
; VI-NEXT: s_add_u32 s21, s21, 3
|
|
; VI-NEXT: s_addc_u32 s20, s20, 0
|
|
; VI-NEXT: s_add_u32 s23, s23, 3
|
|
; VI-NEXT: s_addc_u32 s22, s22, 0
|
|
; VI-NEXT: s_add_u32 s25, s25, 3
|
|
; VI-NEXT: s_addc_u32 s24, s24, 0
|
|
; VI-NEXT: s_add_u32 s40, s40, 3
|
|
; VI-NEXT: s_addc_u32 s26, s26, 0
|
|
; VI-NEXT: s_add_u32 s42, s42, 3
|
|
; VI-NEXT: s_addc_u32 s41, s41, 0
|
|
; VI-NEXT: s_lshr_b32 s27, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s28, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s29, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s43, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s44, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s42, 16
|
|
; VI-NEXT: .LBB41_3: ; %end
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s42
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s41
|
|
; VI-NEXT: s_lshl_b32 s41, s78, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s41
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s40
|
|
; VI-NEXT: s_lshl_b32 s41, s77, 16
|
|
; VI-NEXT: s_or_b32 s40, s40, s41
|
|
; VI-NEXT: s_and_b32 s26, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s41, s76, 16
|
|
; VI-NEXT: s_or_b32 s26, s26, s41
|
|
; VI-NEXT: s_and_b32 s25, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s41, s75, 16
|
|
; VI-NEXT: s_or_b32 s25, s25, s41
|
|
; VI-NEXT: s_and_b32 s24, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s41, s74, 16
|
|
; VI-NEXT: s_or_b32 s24, s24, s41
|
|
; VI-NEXT: s_and_b32 s23, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s41, s73, 16
|
|
; VI-NEXT: s_or_b32 s23, s23, s41
|
|
; VI-NEXT: s_and_b32 s22, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s41, s72, 16
|
|
; VI-NEXT: s_or_b32 s22, s22, s41
|
|
; VI-NEXT: s_and_b32 s21, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s41, s63, 16
|
|
; VI-NEXT: s_or_b32 s21, s21, s41
|
|
; VI-NEXT: s_and_b32 s20, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s41, s62, 16
|
|
; VI-NEXT: s_or_b32 s20, s20, s41
|
|
; VI-NEXT: s_and_b32 s19, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s41, s61, 16
|
|
; VI-NEXT: s_or_b32 s19, s19, s41
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s41, s60, 16
|
|
; VI-NEXT: s_or_b32 s18, s18, s41
|
|
; VI-NEXT: s_and_b32 s17, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s59, 16
|
|
; VI-NEXT: s_or_b32 s17, s17, s41
|
|
; VI-NEXT: s_and_b32 s16, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s41, s58, 16
|
|
; VI-NEXT: s_or_b32 s16, s16, s41
|
|
; VI-NEXT: s_and_b32 s15, 0xffff, s15
|
|
; VI-NEXT: s_lshl_b32 s41, s57, 16
|
|
; VI-NEXT: s_or_b32 s15, s15, s41
|
|
; VI-NEXT: s_and_b32 s14, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s41, s56, 16
|
|
; VI-NEXT: s_or_b32 s14, s14, s41
|
|
; VI-NEXT: s_and_b32 s13, 0xffff, s13
|
|
; VI-NEXT: s_lshl_b32 s41, s47, 16
|
|
; VI-NEXT: s_or_b32 s13, s13, s41
|
|
; VI-NEXT: s_and_b32 s12, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s41, s46, 16
|
|
; VI-NEXT: s_or_b32 s12, s12, s41
|
|
; VI-NEXT: s_and_b32 s11, 0xffff, s11
|
|
; VI-NEXT: s_lshl_b32 s41, s45, 16
|
|
; VI-NEXT: s_or_b32 s11, s11, s41
|
|
; VI-NEXT: s_and_b32 s10, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s41, s44, 16
|
|
; VI-NEXT: s_or_b32 s10, s10, s41
|
|
; VI-NEXT: s_and_b32 s9, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s41, s43, 16
|
|
; VI-NEXT: s_and_b32 s8, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s29, s29, 16
|
|
; VI-NEXT: s_and_b32 s6, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s28, s28, 16
|
|
; VI-NEXT: s_and_b32 s7, 0xffff, s7
|
|
; VI-NEXT: s_lshl_b32 s27, s27, 16
|
|
; VI-NEXT: s_or_b32 s9, s9, s41
|
|
; VI-NEXT: s_or_b32 s8, s8, s29
|
|
; VI-NEXT: s_or_b32 s6, s6, s28
|
|
; VI-NEXT: s_or_b32 s7, s7, s27
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s40
|
|
; VI-NEXT: v_mov_b32_e32 v3, s26
|
|
; VI-NEXT: v_mov_b32_e32 v4, s25
|
|
; VI-NEXT: v_mov_b32_e32 v5, s24
|
|
; VI-NEXT: v_mov_b32_e32 v6, s23
|
|
; VI-NEXT: v_mov_b32_e32 v7, s22
|
|
; VI-NEXT: v_mov_b32_e32 v8, s21
|
|
; VI-NEXT: v_mov_b32_e32 v9, s20
|
|
; VI-NEXT: v_mov_b32_e32 v10, s19
|
|
; VI-NEXT: v_mov_b32_e32 v11, s18
|
|
; VI-NEXT: v_mov_b32_e32 v12, s17
|
|
; VI-NEXT: v_mov_b32_e32 v13, s16
|
|
; VI-NEXT: v_mov_b32_e32 v14, s15
|
|
; VI-NEXT: v_mov_b32_e32 v15, s14
|
|
; VI-NEXT: v_mov_b32_e32 v16, s13
|
|
; VI-NEXT: v_mov_b32_e32 v17, s12
|
|
; VI-NEXT: v_mov_b32_e32 v18, s11
|
|
; VI-NEXT: v_mov_b32_e32 v19, s10
|
|
; VI-NEXT: v_mov_b32_e32 v20, s9
|
|
; VI-NEXT: v_mov_b32_e32 v21, s8
|
|
; VI-NEXT: v_mov_b32_e32 v22, s6
|
|
; VI-NEXT: v_mov_b32_e32 v23, s7
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB41_4:
|
|
; 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: ; implicit-def: $sgpr43
|
|
; VI-NEXT: ; implicit-def: $sgpr29
|
|
; VI-NEXT: ; implicit-def: $sgpr28
|
|
; VI-NEXT: ; implicit-def: $sgpr27
|
|
; VI-NEXT: s_branch .LBB41_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v12i64_to_v48i16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s24
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v11
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s25
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s26
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v13
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s27
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v14
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s28
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v15
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s29
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v17
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v18
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v19
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s17, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s18, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s19, v15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s20, v0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s21, v1
|
|
; GFX9-NEXT: v_readfirstlane_b32 s22, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s23, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s24, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s25, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s26, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s27, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s28, v8
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_readfirstlane_b32 s29, v9
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB41_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s44, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s6, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB41_3
|
|
; GFX9-NEXT: .LBB41_2: ; %cmp.true
|
|
; 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_add_u32 s14, s14, 3
|
|
; GFX9-NEXT: s_addc_u32 s15, s15, 0
|
|
; GFX9-NEXT: s_add_u32 s12, s12, 3
|
|
; GFX9-NEXT: s_addc_u32 s13, s13, 0
|
|
; GFX9-NEXT: s_add_u32 s10, s10, 3
|
|
; GFX9-NEXT: s_addc_u32 s11, s11, 0
|
|
; GFX9-NEXT: s_add_u32 s8, s8, 3
|
|
; GFX9-NEXT: s_addc_u32 s9, s9, 0
|
|
; GFX9-NEXT: s_add_u32 s6, s6, 3
|
|
; GFX9-NEXT: s_addc_u32 s7, s7, 0
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s44, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s6, 16
|
|
; GFX9-NEXT: .LBB41_3: ; %end
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s6, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s5, s7, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s8, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s9, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s10, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s11, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s12, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s13, s72
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s14, s63
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s15, s62
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s16, s61
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s17, s60
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s59
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s58
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s57
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s56
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s47
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s46
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s45
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s44
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s4
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s5
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s27
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB41_4:
|
|
; 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: ; implicit-def: $sgpr43
|
|
; GFX9-NEXT: ; implicit-def: $sgpr42
|
|
; GFX9-NEXT: ; implicit-def: $sgpr41
|
|
; GFX9-NEXT: ; implicit-def: $sgpr40
|
|
; GFX9-NEXT: s_branch .LBB41_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v12i64_to_v48i16_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v7, s0 :: v_dual_mov_b32 v8, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v11, s16 :: v_dual_mov_b32 v12, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v13, s18 :: v_dual_mov_b32 v14, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v15, s20 :: v_dual_mov_b32 v16, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v17, s22 :: v_dual_mov_b32 v18, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v19, s24 :: v_dual_mov_b32 v20, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v21, s26 :: v_dual_mov_b32 v22, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v23, s28 :: v_dual_mov_b32 v24, s29
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s0, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s1, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s2, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s3, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v11
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v12
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v14
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v15
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v16
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v17
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v18
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v19
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v20
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v21
|
|
; GFX11-NEXT: v_readfirstlane_b32 s15, v22
|
|
; GFX11-NEXT: v_readfirstlane_b32 s16, v23
|
|
; GFX11-NEXT: v_readfirstlane_b32 s17, v24
|
|
; GFX11-NEXT: v_readfirstlane_b32 s18, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s19, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s20, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s21, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s23, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s22, v5
|
|
; GFX11-NEXT: s_mov_b32 s74, 0
|
|
; GFX11-NEXT: s_and_b32 s24, vcc_lo, exec_lo
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB41_4
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-NEXT: s_lshr_b32 s24, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s25, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s26, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s27, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s28, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s29, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s15, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s14, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s0, 16
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s74
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB41_3
|
|
; GFX11-NEXT: .LBB41_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_u32 s23, s23, 3
|
|
; GFX11-NEXT: s_addc_u32 s22, s22, 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 s14, s14, 3
|
|
; GFX11-NEXT: s_addc_u32 s15, s15, 0
|
|
; GFX11-NEXT: s_add_u32 s12, s12, 3
|
|
; GFX11-NEXT: s_addc_u32 s13, s13, 0
|
|
; GFX11-NEXT: s_add_u32 s10, s10, 3
|
|
; GFX11-NEXT: s_addc_u32 s11, s11, 0
|
|
; GFX11-NEXT: s_add_u32 s8, s8, 3
|
|
; GFX11-NEXT: s_addc_u32 s9, s9, 0
|
|
; GFX11-NEXT: s_add_u32 s6, s6, 3
|
|
; GFX11-NEXT: s_addc_u32 s7, s7, 0
|
|
; GFX11-NEXT: s_add_u32 s4, s4, 3
|
|
; GFX11-NEXT: s_addc_u32 s5, s5, 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 s24, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s25, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s26, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s27, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s28, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s29, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s15, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s14, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, 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, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s62
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s60
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s59
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s58
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s57
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s56
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s44
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s40
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s29
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s28
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s27
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s26
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s25
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s24
|
|
; 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, s23 :: v_dual_mov_b32 v23, s22
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB41_4:
|
|
; 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: $sgpr29
|
|
; GFX11-NEXT: ; implicit-def: $sgpr28
|
|
; GFX11-NEXT: ; implicit-def: $sgpr27
|
|
; GFX11-NEXT: ; implicit-def: $sgpr26
|
|
; GFX11-NEXT: ; implicit-def: $sgpr25
|
|
; GFX11-NEXT: ; implicit-def: $sgpr24
|
|
; 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 <12 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <12 x i64> %a1 to <48 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x i64> %a to <48 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x i16> %phi
|
|
}
|
|
|
|
define <12 x i64> @bitcast_v48i16_to_v12i64(<48 x i16> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v48i16_to_v12i64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v48, v14
|
|
; SI-NEXT: v_mov_b32_e32 v49, v12
|
|
; SI-NEXT: v_mov_b32_e32 v50, v10
|
|
; SI-NEXT: v_mov_b32_e32 v51, v8
|
|
; SI-NEXT: v_mov_b32_e32 v52, v6
|
|
; SI-NEXT: v_mov_b32_e32 v53, v4
|
|
; SI-NEXT: v_mov_b32_e32 v54, v2
|
|
; SI-NEXT: v_mov_b32_e32 v55, v0
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:68
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:60
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshlrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v43, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v42, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v41, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v29
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v8
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v12
|
|
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:52
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:44
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:40
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:36
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:196 ; 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 v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v55
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v54
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v53
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v52
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v51
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v50
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v49
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v48
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v47
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v39
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v46
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v38
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v45
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v37
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v44
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v36
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; 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: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v43
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v35
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v42
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v34
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v41
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v33
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v40
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v32
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v63
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v62
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v61
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v60
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v59
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v58
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v57
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v56
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; 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 v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v55
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v54
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v53
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v52
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v51
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v50
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v49
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v48
|
|
; 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_or_b32_e32 v0, v47, v0
|
|
; SI-NEXT: s_mov_b32 s6, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v1, v39, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v46, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v38, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v45, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v37, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v44, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v36, v7
|
|
; 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: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; 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_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v8, v43, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v35, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v42, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v34, v11
|
|
; SI-NEXT: v_or_b32_e32 v12, v41, v12
|
|
; SI-NEXT: v_or_b32_e32 v13, v33, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v40, v14
|
|
; SI-NEXT: v_or_b32_e32 v15, v32, v15
|
|
; SI-NEXT: v_or_b32_e32 v16, v63, v16
|
|
; SI-NEXT: v_or_b32_e32 v17, v62, v17
|
|
; SI-NEXT: v_or_b32_e32 v18, v61, v18
|
|
; SI-NEXT: v_or_b32_e32 v19, v60, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v59, v20
|
|
; SI-NEXT: v_or_b32_e32 v21, v58, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v57, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v56, v23
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9
|
|
; 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: v_add_i32_e32 v20, vcc, s6, v20
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v22
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 0x30000, v23
|
|
; 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 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v48i16_to_v12i64:
|
|
; 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: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_mov_b32_e32 v32, v23
|
|
; VI-NEXT: v_mov_b32_e32 v33, v22
|
|
; VI-NEXT: v_mov_b32_e32 v34, v21
|
|
; VI-NEXT: v_mov_b32_e32 v35, v20
|
|
; VI-NEXT: v_mov_b32_e32 v36, v19
|
|
; VI-NEXT: v_mov_b32_e32 v37, v18
|
|
; VI-NEXT: v_mov_b32_e32 v38, v17
|
|
; VI-NEXT: v_mov_b32_e32 v39, v16
|
|
; VI-NEXT: v_mov_b32_e32 v48, v15
|
|
; VI-NEXT: v_mov_b32_e32 v49, v14
|
|
; VI-NEXT: v_mov_b32_e32 v50, v13
|
|
; VI-NEXT: v_mov_b32_e32 v51, v12
|
|
; VI-NEXT: v_mov_b32_e32 v52, v11
|
|
; VI-NEXT: v_mov_b32_e32 v53, v10
|
|
; VI-NEXT: v_mov_b32_e32 v54, v9
|
|
; VI-NEXT: v_mov_b32_e32 v55, v8
|
|
; VI-NEXT: v_mov_b32_e32 v40, v7
|
|
; VI-NEXT: v_mov_b32_e32 v41, v6
|
|
; VI-NEXT: v_mov_b32_e32 v42, v5
|
|
; VI-NEXT: v_mov_b32_e32 v43, v4
|
|
; VI-NEXT: v_mov_b32_e32 v44, v3
|
|
; VI-NEXT: v_mov_b32_e32 v45, v2
|
|
; VI-NEXT: v_mov_b32_e32 v46, v1
|
|
; VI-NEXT: v_mov_b32_e32 v47, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 v23, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v23, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v23, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v23, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v23, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v23, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v23, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v23, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v23, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v23, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v23, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v23, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v23, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v23, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v23, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v23, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v23, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v23, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v23, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v23, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v23, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v23, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v23, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v23, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v44, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v43, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v42, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v41, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v40, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v55, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v53, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v52, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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 v23, 3
|
|
; VI-NEXT: v_add_u16_e32 v0, 3, v47
|
|
; VI-NEXT: v_add_u16_sdwa v1, v47, v23 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_add_u16_e32 v1, 3, v46
|
|
; VI-NEXT: v_add_u16_sdwa v2, v46, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v45
|
|
; VI-NEXT: v_add_u16_sdwa v3, v45, v23 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, v44
|
|
; VI-NEXT: v_add_u16_sdwa v4, v44, v23 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, v43
|
|
; VI-NEXT: v_add_u16_sdwa v5, v43, v23 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, v42
|
|
; VI-NEXT: v_add_u16_sdwa v6, v42, v23 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, v41
|
|
; VI-NEXT: v_add_u16_sdwa v7, v41, v23 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, v40
|
|
; VI-NEXT: v_add_u16_sdwa v8, v40, v23 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, v55
|
|
; VI-NEXT: v_add_u16_sdwa v9, v55, v23 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, v54
|
|
; VI-NEXT: v_add_u16_sdwa v10, v54, v23 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, v53
|
|
; VI-NEXT: v_add_u16_sdwa v11, v53, v23 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, v52
|
|
; VI-NEXT: v_add_u16_sdwa v12, v52, v23 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, v51
|
|
; VI-NEXT: v_add_u16_sdwa v13, v51, v23 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, v50
|
|
; VI-NEXT: v_add_u16_sdwa v14, v50, v23 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, v49
|
|
; VI-NEXT: v_add_u16_sdwa v15, v49, v23 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, v48
|
|
; VI-NEXT: v_add_u16_sdwa v16, v48, v23 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, v39
|
|
; VI-NEXT: v_add_u16_sdwa v17, v39, v23 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, v38
|
|
; VI-NEXT: v_add_u16_sdwa v18, v38, v23 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, v37
|
|
; VI-NEXT: v_add_u16_sdwa v19, v37, v23 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, v36
|
|
; VI-NEXT: v_add_u16_sdwa v20, v36, v23 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, v35
|
|
; VI-NEXT: v_add_u16_sdwa v21, v35, v23 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, v34
|
|
; VI-NEXT: v_add_u16_sdwa v22, v34, v23 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, v33
|
|
; VI-NEXT: v_add_u16_sdwa v24, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v22, v22, v24
|
|
; VI-NEXT: v_add_u16_e32 v24, 3, v32
|
|
; VI-NEXT: v_add_u16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB42_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48i16_to_v12i64:
|
|
; 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 v33, v22
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v13
|
|
; 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, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v12
|
|
; 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, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v11
|
|
; 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, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v10
|
|
; 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, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v1
|
|
; 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, v53
|
|
; 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, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v47
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 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:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; 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: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; 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 v9, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, 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(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, 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: .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_v48i16_to_v12i64:
|
|
; 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, v24
|
|
; 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: .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_v48i16_to_v12i64:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v24
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v53, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v55, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v64, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v52, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v51, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v50, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v49, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v48, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v39, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v38, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v37, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v36, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v35, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v34, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v33, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v32, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v31, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v30, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v29, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v28, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v27, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v26, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v25, v23, 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: .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 <48 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <48 x i16> %a1 to <12 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x i16> %a to <12 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x i64> %phi
|
|
}
|
|
|
|
define inreg <12 x i64> @bitcast_v48i16_to_v12i64_scalar(<48 x i16> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v48i16_to_v12i64_scalar:
|
|
; 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:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v61, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v63, v2
|
|
; SI-NEXT: v_mov_b32_e32 v56, v0
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: v_mov_b32_e32 v31, v22
|
|
; SI-NEXT: v_mov_b32_e32 v34, v20
|
|
; SI-NEXT: v_mov_b32_e32 v35, v18
|
|
; SI-NEXT: v_mov_b32_e32 v36, v16
|
|
; SI-NEXT: v_mov_b32_e32 v37, v14
|
|
; SI-NEXT: v_mov_b32_e32 v38, v12
|
|
; SI-NEXT: v_mov_b32_e32 v39, v10
|
|
; SI-NEXT: v_mov_b32_e32 v48, v8
|
|
; SI-NEXT: v_mov_b32_e32 v50, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v55, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v41, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v52, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v2
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v4
|
|
; SI-NEXT: s_cbranch_scc0 .LBB43_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v56
|
|
; SI-NEXT: v_or_b32_e32 v7, v0, v54
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v61
|
|
; SI-NEXT: v_or_b32_e32 v9, v0, v32
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v50
|
|
; SI-NEXT: v_or_b32_e32 v10, v0, v55
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; SI-NEXT: v_or_b32_e32 v11, v0, v62
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; SI-NEXT: v_or_b32_e32 v12, v0, v46
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; SI-NEXT: v_or_b32_e32 v13, v0, v45
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; SI-NEXT: v_or_b32_e32 v14, v0, v60
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; SI-NEXT: v_or_b32_e32 v15, v0, v59
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s17, 16
|
|
; SI-NEXT: v_or_b32_e32 v16, v0, v58
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; SI-NEXT: s_or_b32 s4, s4, s5
|
|
; SI-NEXT: s_and_b32 s5, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s19, 16
|
|
; SI-NEXT: v_or_b32_e32 v17, v0, v41
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v31
|
|
; SI-NEXT: s_or_b32 s5, s5, s6
|
|
; SI-NEXT: s_and_b32 s6, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s21, 16
|
|
; SI-NEXT: v_or_b32_e32 v18, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v24
|
|
; SI-NEXT: s_or_b32 s6, s6, s7
|
|
; SI-NEXT: s_and_b32 s7, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s23, 16
|
|
; SI-NEXT: v_or_b32_e32 v19, v0, v52
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v26
|
|
; SI-NEXT: s_or_b32 s7, s7, s8
|
|
; SI-NEXT: s_and_b32 s8, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s25, 16
|
|
; SI-NEXT: v_or_b32_e32 v20, v0, v51
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v28
|
|
; SI-NEXT: s_or_b32 s8, s8, s9
|
|
; SI-NEXT: s_and_b32 s9, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s10, s27, 16
|
|
; SI-NEXT: v_or_b32_e32 v21, v0, v29
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v30
|
|
; SI-NEXT: s_or_b32 s9, s9, s10
|
|
; SI-NEXT: s_and_b32 s10, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s11, s29, 16
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v63
|
|
; SI-NEXT: v_or_b32_e32 v22, v0, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; SI-NEXT: s_or_b32 s10, s10, s11
|
|
; SI-NEXT: v_or_b32_e32 v8, v1, v57
|
|
; SI-NEXT: v_or_b32_e32 v23, v0, v25
|
|
; SI-NEXT: v_mov_b32_e32 v0, s4
|
|
; SI-NEXT: v_mov_b32_e32 v1, s5
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: v_mov_b32_e32 v3, s7
|
|
; SI-NEXT: v_mov_b32_e32 v4, s8
|
|
; SI-NEXT: v_mov_b32_e32 v5, s9
|
|
; SI-NEXT: v_mov_b32_e32 v6, s10
|
|
; SI-NEXT: s_cbranch_execnz .LBB43_3
|
|
; SI-NEXT: .LBB43_2: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v56
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v54, v0
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v61
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v32, v0
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v50
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_or_b32_e32 v0, v55, v0
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v48
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v62, v0
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v39
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v0, v46, v0
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v38
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v45, v0
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v37
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v60, v0
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v36
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v59, v0
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v35
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v58, v0
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v34
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v41, v0
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v31
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v40, v0
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v24
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v52, v0
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v26
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v51, v0
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s17, 16
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v28
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_and_b32 s5, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s19, 16
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_or_b32 s5, s6, s5
|
|
; SI-NEXT: s_and_b32 s6, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s21, 16
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v29, v0
|
|
; SI-NEXT: s_or_b32 s6, s7, s6
|
|
; SI-NEXT: s_and_b32 s7, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s23, 16
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v30
|
|
; SI-NEXT: s_or_b32 s7, s8, s7
|
|
; SI-NEXT: s_and_b32 s8, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s25, 16
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_or_b32 s8, s9, s8
|
|
; SI-NEXT: s_and_b32 s9, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s10, s27, 16
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v27, v0
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v63
|
|
; SI-NEXT: s_or_b32 s9, s10, s9
|
|
; SI-NEXT: s_and_b32 s10, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s11, s29, 16
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v49
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: s_or_b32 s10, s11, s10
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v57, v1
|
|
; SI-NEXT: s_add_i32 s4, s4, 0x30000
|
|
; SI-NEXT: s_add_i32 s5, s5, 0x30000
|
|
; SI-NEXT: s_add_i32 s6, s6, 0x30000
|
|
; SI-NEXT: s_add_i32 s7, s7, 0x30000
|
|
; SI-NEXT: s_add_i32 s8, s8, 0x30000
|
|
; SI-NEXT: s_add_i32 s9, s9, 0x30000
|
|
; SI-NEXT: s_add_i32 s10, s10, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v0, v25, v0
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 0x30000, v1
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 0x30000, v0
|
|
; SI-NEXT: v_mov_b32_e32 v0, s4
|
|
; SI-NEXT: v_mov_b32_e32 v1, s5
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: v_mov_b32_e32 v3, s7
|
|
; SI-NEXT: v_mov_b32_e32 v4, s8
|
|
; SI-NEXT: v_mov_b32_e32 v5, s9
|
|
; SI-NEXT: v_mov_b32_e32 v6, s10
|
|
; SI-NEXT: .LBB43_3: ; %end
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB43_4:
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v46, v51
|
|
; SI-NEXT: v_mov_b32_e32 v51, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v34
|
|
; SI-NEXT: v_mov_b32_e32 v34, v30
|
|
; SI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_mov_b32_e32 v47, v52
|
|
; SI-NEXT: v_mov_b32_e32 v52, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v35
|
|
; SI-NEXT: v_mov_b32_e32 v35, v28
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v55, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v41
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v41, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v26
|
|
; SI-NEXT: v_mov_b32_e32 v42, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v24
|
|
; SI-NEXT: v_mov_b32_e32 v33, v32
|
|
; SI-NEXT: v_mov_b32_e32 v32, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v40
|
|
; SI-NEXT: v_mov_b32_e32 v40, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v31
|
|
; SI-NEXT: v_mov_b32_e32 v43, v25
|
|
; SI-NEXT: v_mov_b32_e32 v44, v27
|
|
; SI-NEXT: v_mov_b32_e32 v53, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v29
|
|
; 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: v_mov_b32_e32 v26, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v55
|
|
; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v30, v34
|
|
; SI-NEXT: v_mov_b32_e32 v34, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v51
|
|
; SI-NEXT: v_mov_b32_e32 v51, v46
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v29, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v53
|
|
; SI-NEXT: v_mov_b32_e32 v27, v44
|
|
; SI-NEXT: v_mov_b32_e32 v25, v43
|
|
; SI-NEXT: v_mov_b32_e32 v31, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v40
|
|
; SI-NEXT: v_mov_b32_e32 v40, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v32
|
|
; SI-NEXT: v_mov_b32_e32 v32, v33
|
|
; SI-NEXT: v_mov_b32_e32 v24, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v42
|
|
; SI-NEXT: v_mov_b32_e32 v28, v35
|
|
; SI-NEXT: v_mov_b32_e32 v35, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v52
|
|
; SI-NEXT: v_mov_b32_e32 v52, v47
|
|
; SI-NEXT: s_branch .LBB43_2
|
|
;
|
|
; VI-LABEL: bitcast_v48i16_to_v12i64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v32, v9
|
|
; VI-NEXT: v_mov_b32_e32 v33, v8
|
|
; VI-NEXT: v_mov_b32_e32 v34, v7
|
|
; VI-NEXT: v_mov_b32_e32 v35, v6
|
|
; VI-NEXT: v_mov_b32_e32 v36, v5
|
|
; VI-NEXT: v_mov_b32_e32 v37, v4
|
|
; VI-NEXT: v_mov_b32_e32 v38, v3
|
|
; VI-NEXT: v_mov_b32_e32 v39, v2
|
|
; VI-NEXT: v_mov_b32_e32 v48, v1
|
|
; VI-NEXT: v_mov_b32_e32 v49, v0
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; 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, s43, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s44, s42, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s44
|
|
; VI-NEXT: s_and_b32 s44, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s45, s41, 16
|
|
; VI-NEXT: s_or_b32 s44, s44, s45
|
|
; VI-NEXT: s_and_b32 s45, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s46, s40, 16
|
|
; VI-NEXT: s_or_b32 s45, s45, s46
|
|
; VI-NEXT: s_and_b32 s46, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s47, s15, 16
|
|
; VI-NEXT: s_or_b32 s46, s46, s47
|
|
; VI-NEXT: s_and_b32 s47, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s56, s14, 16
|
|
; VI-NEXT: v_mov_b32_e32 v0, 16
|
|
; VI-NEXT: s_or_b32 s47, s47, s56
|
|
; VI-NEXT: s_and_b32 s56, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s57, s13, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s56, s56, s57
|
|
; VI-NEXT: s_and_b32 s57, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s58, s12, 16
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s57, s57, s58
|
|
; VI-NEXT: s_and_b32 s58, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s59, s11, 16
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s58, s58, s59
|
|
; VI-NEXT: s_and_b32 s59, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s60, s10, 16
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s59, s59, s60
|
|
; VI-NEXT: s_and_b32 s60, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s61, s9, 16
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s60, s60, s61
|
|
; VI-NEXT: s_and_b32 s61, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s62, s8, 16
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s61, s61, s62
|
|
; VI-NEXT: s_and_b32 s62, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s63, s7, 16
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s62, s62, s63
|
|
; VI-NEXT: s_and_b32 s63, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s72, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s63, s63, s72
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v0, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s44
|
|
; VI-NEXT: v_mov_b32_e32 v3, s45
|
|
; VI-NEXT: v_mov_b32_e32 v4, s46
|
|
; VI-NEXT: v_mov_b32_e32 v5, s47
|
|
; VI-NEXT: v_mov_b32_e32 v6, s56
|
|
; VI-NEXT: v_mov_b32_e32 v7, s57
|
|
; VI-NEXT: v_mov_b32_e32 v8, s58
|
|
; VI-NEXT: v_mov_b32_e32 v9, s59
|
|
; VI-NEXT: v_mov_b32_e32 v10, s60
|
|
; VI-NEXT: v_mov_b32_e32 v11, s61
|
|
; VI-NEXT: v_mov_b32_e32 v12, s62
|
|
; VI-NEXT: v_mov_b32_e32 v13, s63
|
|
; VI-NEXT: s_cbranch_execnz .LBB43_3
|
|
; VI-NEXT: .LBB43_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v1, 16
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v49
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v48
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s43, 16
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: v_add_u32_e32 v15, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v39
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_and_b32 s5, s17, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s16, s42, 16
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s5, s16, s5
|
|
; VI-NEXT: s_and_b32 s16, s18, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s17, s41, 16
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v38
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s16, s17, s16
|
|
; VI-NEXT: s_and_b32 s17, s19, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s18, s40, 16
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s17, s18, s17
|
|
; VI-NEXT: s_and_b32 s18, s20, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s15, s15, 16
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: v_add_u32_e32 v17, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v37
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s15, s15, s18
|
|
; VI-NEXT: s_and_b32 s18, s21, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s14, s14, 16
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s14, s14, s18
|
|
; VI-NEXT: s_and_b32 s18, s22, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s13, s13, 16
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v36
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s13, s13, s18
|
|
; VI-NEXT: s_and_b32 s18, s23, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s12, s12, 16
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s12, s12, s18
|
|
; VI-NEXT: s_and_b32 s18, s24, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s11, s11, 16
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: v_add_u32_e32 v19, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v35
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s11, s11, s18
|
|
; VI-NEXT: s_and_b32 s18, s25, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s10, s10, 16
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s10, s10, s18
|
|
; VI-NEXT: s_and_b32 s18, s26, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s9, s9, 16
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v34
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s9, s9, s18
|
|
; VI-NEXT: s_and_b32 s18, s27, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s8, s8, 16
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s8, s8, s18
|
|
; VI-NEXT: s_and_b32 s18, s28, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s7, s7, 16
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: v_add_u32_e32 v21, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v33
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s7, s7, s18
|
|
; VI-NEXT: s_and_b32 s18, s29, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s6, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s6, s6, s18
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v32
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v1, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_add_i32 s4, s4, 0x30000
|
|
; VI-NEXT: s_add_i32 s5, s5, 0x30000
|
|
; VI-NEXT: s_add_i32 s16, s16, 0x30000
|
|
; VI-NEXT: s_add_i32 s17, s17, 0x30000
|
|
; VI-NEXT: s_add_i32 s15, s15, 0x30000
|
|
; VI-NEXT: s_add_i32 s14, s14, 0x30000
|
|
; VI-NEXT: s_add_i32 s13, s13, 0x30000
|
|
; VI-NEXT: s_add_i32 s12, s12, 0x30000
|
|
; VI-NEXT: s_add_i32 s11, s11, 0x30000
|
|
; VI-NEXT: s_add_i32 s10, s10, 0x30000
|
|
; VI-NEXT: s_add_i32 s9, s9, 0x30000
|
|
; VI-NEXT: s_add_i32 s8, s8, 0x30000
|
|
; VI-NEXT: s_add_i32 s7, s7, 0x30000
|
|
; VI-NEXT: s_add_i32 s6, s6, 0x30000
|
|
; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: v_add_u32_e32 v23, vcc, 0x30000, v0
|
|
; 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, s15
|
|
; VI-NEXT: v_mov_b32_e32 v5, s14
|
|
; VI-NEXT: v_mov_b32_e32 v6, s13
|
|
; VI-NEXT: v_mov_b32_e32 v7, s12
|
|
; VI-NEXT: v_mov_b32_e32 v8, s11
|
|
; VI-NEXT: v_mov_b32_e32 v9, s10
|
|
; VI-NEXT: v_mov_b32_e32 v10, s9
|
|
; VI-NEXT: v_mov_b32_e32 v11, s8
|
|
; VI-NEXT: v_mov_b32_e32 v12, s7
|
|
; VI-NEXT: v_mov_b32_e32 v13, s6
|
|
; VI-NEXT: .LBB43_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB43_4:
|
|
; 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_branch .LBB43_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v48i16_to_v12i64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v0
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, 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: buffer_store_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v32
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v33
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v49
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s17, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s18, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s19, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s20, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s21, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s22, s12
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s23, s13
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s24, s14
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s25, s15
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s29, s40
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB43_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v33
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s19
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB43_3
|
|
; GFX9-NEXT: .LBB43_2: ; %cmp.true
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v49
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v48
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v39
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v38
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v37
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v36
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v35
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v34
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v33
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v23
|
|
; GFX9-NEXT: v_pk_add_u16 v0, s6, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, s7, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, s8, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, s9, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, s10, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, s11, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, s12, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, s13, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, s14, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v9, s15, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, s16, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, s17, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, s18, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, s19, 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: .LBB43_3: ; %end
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB43_4:
|
|
; 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_branch .LBB43_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v12i64_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB43_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; 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, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB43_3
|
|
; GFX11-TRUE16-NEXT: .LBB43_2: ; %cmp.true
|
|
; 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, s4, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s17, 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: .LBB43_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB43_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB43_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v48i16_to_v12i64_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s4, s0, s4
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s44
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s16, s5
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s17, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s18, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s19, s8
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s20, s9
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s21, s10
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s22, s11
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s23, s12
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s24, s13
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s25, s14
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s26, s43
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB43_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB43_3
|
|
; GFX11-FAKE16-NEXT: .LBB43_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, s4, 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 v2, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, s17, 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: .LBB43_3: ; %end
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB43_4:
|
|
; GFX11-FAKE16-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
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB43_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <48 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <48 x i16> %a1 to <12 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x i16> %a to <12 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x i64> %phi
|
|
}
|
|
|
|
define <48 x half> @bitcast_v12i64_to_v48f16(<12 x i64> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v12i64_to_v48f16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v25
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; 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: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; 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: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; 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_lshrrev_b32_e32 v25, 16, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v1
|
|
; SI-NEXT: ; implicit-def: $vgpr1
|
|
; SI-NEXT: ; implicit-def: $vgpr2
|
|
; SI-NEXT: ; implicit-def: $vgpr3
|
|
; SI-NEXT: ; implicit-def: $vgpr4
|
|
; SI-NEXT: ; implicit-def: $vgpr5
|
|
; SI-NEXT: ; implicit-def: $vgpr6
|
|
; 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: $vgpr13
|
|
; SI-NEXT: ; implicit-def: $vgpr14
|
|
; SI-NEXT: ; implicit-def: $vgpr15
|
|
; SI-NEXT: ; implicit-def: $vgpr16
|
|
; SI-NEXT: ; implicit-def: $vgpr17
|
|
; SI-NEXT: ; implicit-def: $vgpr18
|
|
; SI-NEXT: ; implicit-def: $vgpr19
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr22
|
|
; SI-NEXT: ; implicit-def: $vgpr23
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v1, vcc, 3, v1
|
|
; SI-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
|
|
; SI-NEXT: v_addc_u32_e32 v4, vcc, 0, v4, vcc
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; SI-NEXT: v_addc_u32_e32 v6, vcc, 0, v6, vcc
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
|
|
; SI-NEXT: v_addc_u32_e32 v8, vcc, 0, v8, vcc
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: v_addc_u32_e32 v10, vcc, 0, v10, vcc
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: v_addc_u32_e32 v12, vcc, 0, v12, vcc
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: v_addc_u32_e32 v14, vcc, 0, v14, vcc
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_addc_u32_e32 v16, vcc, 0, v16, vcc
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_addc_u32_e32 v18, vcc, 0, v18, vcc
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_addc_u32_e32 v20, vcc, 0, v20, vcc
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_addc_u32_e32 v22, vcc, 0, v22, vcc
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: v_addc_u32_e32 v24, vcc, 0, v24, vcc
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v1
|
|
; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v3
|
|
; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v5
|
|
; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v7
|
|
; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v9
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v11
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v13
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v15
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v51
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v49
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v36
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v48
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v52
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v40
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v45
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v47
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v61
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v33
|
|
; SI-NEXT: v_mov_b32_e32 v35, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: .LBB44_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v30
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v3
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
|
|
; SI-NEXT: v_or_b32_e32 v2, v4, v2
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v29
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v28
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v26
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v63
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v25
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v61
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v62
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v59
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v60
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v57
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v58
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v56
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v46
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v43
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v44
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v42
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v41
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v55
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v50
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v53
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v48
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v38
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v36
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v34
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v37
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v39
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v49
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v51
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; 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_v12i64_to_v48f16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; VI-NEXT: .LBB44_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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_lshlrev_b32_e32 v27, 16, v27
|
|
; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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_or_b32_sdwa v20, v20, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v12i64_to_v48f16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v54, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v53, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v52, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v51, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v50, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v49, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v48, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v39, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v37, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v36, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v35, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v34, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v33, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v32, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v31, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v30, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v29, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v28, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v27, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v26, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v25, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s4
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v12i64_to_v48f16:
|
|
; 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, v24
|
|
; 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 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_v12i64_to_v48f16:
|
|
; 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, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v53, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v52, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v51, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v50, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v49, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v48, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v39, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v37, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v35, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v33, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v32, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v31, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v30, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v29, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v28, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v27, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v26, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v25, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v24, v23, 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 <12 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <12 x i64> %a1 to <48 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x i64> %a to <48 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x half> %phi
|
|
}
|
|
|
|
define inreg <48 x half> @bitcast_v12i64_to_v48f16_scalar(<12 x i64> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v12i64_to_v48f16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v12, s16
|
|
; SI-NEXT: v_mov_b32_e32 v13, s17
|
|
; SI-NEXT: v_mov_b32_e32 v14, s18
|
|
; SI-NEXT: v_mov_b32_e32 v15, s19
|
|
; SI-NEXT: v_mov_b32_e32 v16, s20
|
|
; SI-NEXT: v_mov_b32_e32 v17, s21
|
|
; SI-NEXT: v_mov_b32_e32 v18, s22
|
|
; SI-NEXT: v_mov_b32_e32 v19, s23
|
|
; SI-NEXT: v_readfirstlane_b32 s22, v12
|
|
; SI-NEXT: v_mov_b32_e32 v12, s24
|
|
; SI-NEXT: v_readfirstlane_b32 s40, v13
|
|
; SI-NEXT: v_mov_b32_e32 v13, s25
|
|
; SI-NEXT: v_readfirstlane_b32 s23, v14
|
|
; SI-NEXT: v_mov_b32_e32 v14, s26
|
|
; SI-NEXT: v_readfirstlane_b32 s41, v15
|
|
; SI-NEXT: v_mov_b32_e32 v15, s27
|
|
; SI-NEXT: v_readfirstlane_b32 s24, v16
|
|
; SI-NEXT: v_mov_b32_e32 v16, s28
|
|
; SI-NEXT: v_readfirstlane_b32 s27, v17
|
|
; SI-NEXT: v_mov_b32_e32 v17, s29
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
|
|
; SI-NEXT: v_readfirstlane_b32 s25, v18
|
|
; SI-NEXT: v_readfirstlane_b32 s26, v19
|
|
; SI-NEXT: v_readfirstlane_b32 s20, v12
|
|
; SI-NEXT: v_readfirstlane_b32 s21, v13
|
|
; SI-NEXT: v_readfirstlane_b32 s18, v14
|
|
; SI-NEXT: v_readfirstlane_b32 s19, v15
|
|
; SI-NEXT: v_readfirstlane_b32 s16, v16
|
|
; SI-NEXT: v_readfirstlane_b32 s17, v17
|
|
; SI-NEXT: v_readfirstlane_b32 s14, v1
|
|
; SI-NEXT: v_readfirstlane_b32 s15, v2
|
|
; SI-NEXT: v_readfirstlane_b32 s12, v3
|
|
; SI-NEXT: v_readfirstlane_b32 s13, v4
|
|
; SI-NEXT: v_readfirstlane_b32 s10, v5
|
|
; SI-NEXT: v_readfirstlane_b32 s11, v6
|
|
; SI-NEXT: v_readfirstlane_b32 s7, v7
|
|
; SI-NEXT: v_readfirstlane_b32 s8, v8
|
|
; SI-NEXT: v_readfirstlane_b32 s6, v9
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: v_readfirstlane_b32 s9, v10
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_cbranch_scc0 .LBB45_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_lshr_b32 s4, s9, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s6, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s8, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s7, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s11, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s10, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s13, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s12, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s15, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s14, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s17, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s16, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s19, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s18, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s21, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s20, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s26, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s25, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s27, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s24, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s41, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s23, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s40, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, s4
|
|
; SI-NEXT: s_lshr_b32 s4, s22, 16
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, s4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, s18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, s21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, s20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, s26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, s25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, s27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, s24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, s41
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, s23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, s40
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, s22
|
|
; SI-NEXT: s_cbranch_execnz .LBB45_3
|
|
; SI-NEXT: .LBB45_2: ; %cmp.true
|
|
; SI-NEXT: s_add_u32 s4, s22, 3
|
|
; SI-NEXT: s_addc_u32 s5, s40, 0
|
|
; SI-NEXT: s_lshr_b32 s22, s4, 16
|
|
; SI-NEXT: s_lshr_b32 s28, s5, 16
|
|
; SI-NEXT: s_add_u32 s23, s23, 3
|
|
; SI-NEXT: s_addc_u32 s29, s41, 0
|
|
; SI-NEXT: s_lshr_b32 s40, s23, 16
|
|
; SI-NEXT: s_lshr_b32 s41, s29, 16
|
|
; SI-NEXT: s_add_u32 s24, s24, 3
|
|
; SI-NEXT: s_addc_u32 s27, s27, 0
|
|
; SI-NEXT: s_lshr_b32 s42, s24, 16
|
|
; SI-NEXT: s_lshr_b32 s43, s27, 16
|
|
; SI-NEXT: s_add_u32 s25, s25, 3
|
|
; SI-NEXT: s_addc_u32 s26, s26, 0
|
|
; SI-NEXT: s_lshr_b32 s44, s25, 16
|
|
; SI-NEXT: s_lshr_b32 s45, s26, 16
|
|
; SI-NEXT: s_add_u32 s20, s20, 3
|
|
; SI-NEXT: s_addc_u32 s21, s21, 0
|
|
; SI-NEXT: s_lshr_b32 s46, s20, 16
|
|
; SI-NEXT: s_lshr_b32 s47, s21, 16
|
|
; SI-NEXT: s_add_u32 s18, s18, 3
|
|
; SI-NEXT: s_addc_u32 s19, s19, 0
|
|
; SI-NEXT: s_lshr_b32 s56, s18, 16
|
|
; SI-NEXT: s_lshr_b32 s57, s19, 16
|
|
; SI-NEXT: s_add_u32 s16, s16, 3
|
|
; SI-NEXT: s_addc_u32 s17, s17, 0
|
|
; SI-NEXT: s_lshr_b32 s58, s16, 16
|
|
; SI-NEXT: s_lshr_b32 s59, s17, 16
|
|
; SI-NEXT: s_add_u32 s14, s14, 3
|
|
; SI-NEXT: s_addc_u32 s15, s15, 0
|
|
; SI-NEXT: s_lshr_b32 s60, s14, 16
|
|
; SI-NEXT: s_lshr_b32 s61, s15, 16
|
|
; SI-NEXT: s_add_u32 s12, s12, 3
|
|
; SI-NEXT: s_addc_u32 s13, s13, 0
|
|
; SI-NEXT: s_lshr_b32 s62, s12, 16
|
|
; SI-NEXT: s_lshr_b32 s63, s13, 16
|
|
; SI-NEXT: s_add_u32 s10, s10, 3
|
|
; SI-NEXT: s_addc_u32 s11, s11, 0
|
|
; SI-NEXT: s_lshr_b32 s72, s10, 16
|
|
; SI-NEXT: s_lshr_b32 s73, s11, 16
|
|
; SI-NEXT: s_add_u32 s7, s7, 3
|
|
; SI-NEXT: s_addc_u32 s8, s8, 0
|
|
; SI-NEXT: s_lshr_b32 s74, s7, 16
|
|
; SI-NEXT: s_lshr_b32 s75, s8, 16
|
|
; SI-NEXT: s_add_u32 s6, s6, 3
|
|
; SI-NEXT: s_addc_u32 s9, s9, 0
|
|
; SI-NEXT: s_lshr_b32 s76, s6, 16
|
|
; SI-NEXT: s_lshr_b32 s77, s9, 16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, s9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, s6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, s8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, s7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, s11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, s10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, s13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, s12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, s15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, s14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, s17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, s19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, s18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, s21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, s20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, s26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, s25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, s27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, s24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, s29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, s23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, s5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, s4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, s77
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, s76
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, s75
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, s74
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, s73
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, s72
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, s63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, s62
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, s61
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, s60
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, s59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, s58
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, s57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, s56
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, s47
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s46
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, s45
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, s44
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, s43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, s42
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, s41
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, s40
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, s28
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, s22
|
|
; SI-NEXT: .LBB45_3: ; %end
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v40, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v55
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v53
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, v51
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v49
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v39
|
|
; SI-NEXT: v_or_b32_e32 v55, v55, v40
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v48
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v37
|
|
; SI-NEXT: buffer_store_dword v55, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v55, vcc, 4, v0
|
|
; SI-NEXT: v_or_b32_e32 v53, v53, v54
|
|
; SI-NEXT: v_lshlrev_b32_e32 v52, 16, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v38
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, v35
|
|
; SI-NEXT: buffer_store_dword v53, v55, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v53, vcc, 8, v0
|
|
; SI-NEXT: v_or_b32_e32 v51, v51, v52
|
|
; SI-NEXT: v_lshlrev_b32_e32 v50, 16, v50
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v36
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v33
|
|
; SI-NEXT: buffer_store_dword v51, v53, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v51, vcc, 12, v0
|
|
; SI-NEXT: v_or_b32_e32 v49, v49, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v39
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, v34
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
|
|
; SI-NEXT: buffer_store_dword v49, v51, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v49, vcc, 16, v0
|
|
; SI-NEXT: v_or_b32_e32 v39, v48, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v37
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: buffer_store_dword v39, v49, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v39, vcc, 20, v0
|
|
; SI-NEXT: v_or_b32_e32 v37, v38, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v35
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v30, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
|
|
; SI-NEXT: buffer_store_dword v37, v39, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v37, vcc, 24, v0
|
|
; SI-NEXT: v_or_b32_e32 v35, v36, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: buffer_store_dword v35, v37, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v35, vcc, 28, v0
|
|
; SI-NEXT: v_or_b32_e32 v33, v34, v33
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: buffer_store_dword v33, v35, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v33, vcc, 32, v0
|
|
; SI-NEXT: v_or_b32_e32 v31, v32, v31
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: buffer_store_dword v31, v33, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v31, vcc, 36, v0
|
|
; SI-NEXT: v_or_b32_e32 v29, v30, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
|
|
; SI-NEXT: buffer_store_dword v29, v31, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v29, vcc, 40, v0
|
|
; SI-NEXT: v_or_b32_e32 v27, v28, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; SI-NEXT: buffer_store_dword v27, v29, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 44, v0
|
|
; SI-NEXT: v_or_b32_e32 v24, v26, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: buffer_store_dword v24, v27, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v25
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: buffer_store_dword v22, v25, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v23
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 52, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v20
|
|
; SI-NEXT: v_or_b32_e32 v20, v22, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: buffer_store_dword v20, v23, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v21
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 56, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: buffer_store_dword v18, v21, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v19
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 60, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v18, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v17
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 64, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v16, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: buffer_store_dword v14, v17, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v15
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 0x44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: buffer_store_dword v12, v15, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v13
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 0x48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v12, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: buffer_store_dword v10, v13, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v11
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 0x4c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: v_or_b32_e32 v7, v10, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: buffer_store_dword v7, v11, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v9
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 0x50, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; SI-NEXT: v_or_b32_e32 v5, v7, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: buffer_store_dword v5, v9, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v8
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 0x54, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v3, v5, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v6
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 0x58, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: buffer_store_dword v2, v5, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v4
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB45_4:
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr22
|
|
; SI-NEXT: ; implicit-def: $vgpr23
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: ; implicit-def: $vgpr18
|
|
; SI-NEXT: ; implicit-def: $vgpr19
|
|
; SI-NEXT: ; implicit-def: $vgpr16
|
|
; SI-NEXT: ; implicit-def: $vgpr17
|
|
; SI-NEXT: ; implicit-def: $vgpr14
|
|
; SI-NEXT: ; implicit-def: $vgpr15
|
|
; SI-NEXT: ; implicit-def: $vgpr12
|
|
; SI-NEXT: ; implicit-def: $vgpr13
|
|
; SI-NEXT: ; implicit-def: $vgpr10
|
|
; SI-NEXT: ; implicit-def: $vgpr11
|
|
; SI-NEXT: ; implicit-def: $vgpr7
|
|
; SI-NEXT: ; implicit-def: $vgpr9
|
|
; SI-NEXT: ; implicit-def: $vgpr5
|
|
; SI-NEXT: ; implicit-def: $vgpr8
|
|
; SI-NEXT: ; implicit-def: $vgpr3
|
|
; SI-NEXT: ; implicit-def: $vgpr6
|
|
; SI-NEXT: ; implicit-def: $vgpr2
|
|
; SI-NEXT: ; implicit-def: $vgpr4
|
|
; SI-NEXT: ; implicit-def: $vgpr1
|
|
; SI-NEXT: s_branch .LBB45_2
|
|
;
|
|
; VI-LABEL: bitcast_v12i64_to_v48f16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_mov_b32_e32 v11, s16
|
|
; VI-NEXT: v_mov_b32_e32 v12, s17
|
|
; VI-NEXT: v_mov_b32_e32 v13, s18
|
|
; VI-NEXT: v_mov_b32_e32 v14, s19
|
|
; VI-NEXT: v_mov_b32_e32 v15, s20
|
|
; VI-NEXT: v_mov_b32_e32 v16, s21
|
|
; VI-NEXT: v_mov_b32_e32 v17, s22
|
|
; VI-NEXT: v_mov_b32_e32 v18, s23
|
|
; VI-NEXT: v_mov_b32_e32 v19, s24
|
|
; VI-NEXT: v_readfirstlane_b32 s42, v11
|
|
; VI-NEXT: v_mov_b32_e32 v11, s25
|
|
; VI-NEXT: v_readfirstlane_b32 s41, v12
|
|
; VI-NEXT: v_mov_b32_e32 v12, s26
|
|
; VI-NEXT: v_readfirstlane_b32 s40, v13
|
|
; VI-NEXT: v_mov_b32_e32 v13, s27
|
|
; VI-NEXT: v_readfirstlane_b32 s26, v14
|
|
; VI-NEXT: v_mov_b32_e32 v14, s28
|
|
; VI-NEXT: v_readfirstlane_b32 s25, v15
|
|
; VI-NEXT: v_mov_b32_e32 v15, s29
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_readfirstlane_b32 s24, v16
|
|
; VI-NEXT: v_readfirstlane_b32 s23, v17
|
|
; VI-NEXT: v_readfirstlane_b32 s22, v18
|
|
; VI-NEXT: v_readfirstlane_b32 s21, v19
|
|
; VI-NEXT: v_readfirstlane_b32 s20, v11
|
|
; VI-NEXT: v_readfirstlane_b32 s19, v12
|
|
; VI-NEXT: v_readfirstlane_b32 s18, v13
|
|
; VI-NEXT: v_readfirstlane_b32 s17, v14
|
|
; VI-NEXT: v_readfirstlane_b32 s16, v15
|
|
; VI-NEXT: v_readfirstlane_b32 s15, v0
|
|
; VI-NEXT: v_readfirstlane_b32 s14, v1
|
|
; VI-NEXT: v_readfirstlane_b32 s13, v2
|
|
; VI-NEXT: v_readfirstlane_b32 s12, v3
|
|
; VI-NEXT: v_readfirstlane_b32 s11, v4
|
|
; VI-NEXT: v_readfirstlane_b32 s10, v5
|
|
; VI-NEXT: v_readfirstlane_b32 s9, v6
|
|
; VI-NEXT: v_readfirstlane_b32 s8, v7
|
|
; VI-NEXT: v_readfirstlane_b32 s6, v8
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_readfirstlane_b32 s7, v9
|
|
; VI-NEXT: s_cbranch_scc0 .LBB45_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_lshr_b32 s27, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s28, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s29, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s43, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s44, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s42, 16
|
|
; VI-NEXT: s_cbranch_execnz .LBB45_3
|
|
; VI-NEXT: .LBB45_2: ; %cmp.true
|
|
; VI-NEXT: s_add_u32 s6, s6, 3
|
|
; VI-NEXT: s_addc_u32 s7, s7, 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 s17, s17, 3
|
|
; VI-NEXT: s_addc_u32 s16, s16, 0
|
|
; VI-NEXT: s_add_u32 s19, s19, 3
|
|
; VI-NEXT: s_addc_u32 s18, s18, 0
|
|
; VI-NEXT: s_add_u32 s21, s21, 3
|
|
; VI-NEXT: s_addc_u32 s20, s20, 0
|
|
; VI-NEXT: s_add_u32 s23, s23, 3
|
|
; VI-NEXT: s_addc_u32 s22, s22, 0
|
|
; VI-NEXT: s_add_u32 s25, s25, 3
|
|
; VI-NEXT: s_addc_u32 s24, s24, 0
|
|
; VI-NEXT: s_add_u32 s40, s40, 3
|
|
; VI-NEXT: s_addc_u32 s26, s26, 0
|
|
; VI-NEXT: s_add_u32 s42, s42, 3
|
|
; VI-NEXT: s_addc_u32 s41, s41, 0
|
|
; VI-NEXT: s_lshr_b32 s27, s7, 16
|
|
; VI-NEXT: s_lshr_b32 s28, s6, 16
|
|
; VI-NEXT: s_lshr_b32 s29, s8, 16
|
|
; VI-NEXT: s_lshr_b32 s43, s9, 16
|
|
; VI-NEXT: s_lshr_b32 s44, s10, 16
|
|
; VI-NEXT: s_lshr_b32 s45, s11, 16
|
|
; VI-NEXT: s_lshr_b32 s46, s12, 16
|
|
; VI-NEXT: s_lshr_b32 s47, s13, 16
|
|
; VI-NEXT: s_lshr_b32 s56, s14, 16
|
|
; VI-NEXT: s_lshr_b32 s57, s15, 16
|
|
; VI-NEXT: s_lshr_b32 s58, s16, 16
|
|
; VI-NEXT: s_lshr_b32 s59, s17, 16
|
|
; VI-NEXT: s_lshr_b32 s60, s18, 16
|
|
; VI-NEXT: s_lshr_b32 s61, s19, 16
|
|
; VI-NEXT: s_lshr_b32 s62, s20, 16
|
|
; VI-NEXT: s_lshr_b32 s63, s21, 16
|
|
; VI-NEXT: s_lshr_b32 s72, s22, 16
|
|
; VI-NEXT: s_lshr_b32 s73, s23, 16
|
|
; VI-NEXT: s_lshr_b32 s74, s24, 16
|
|
; VI-NEXT: s_lshr_b32 s75, s25, 16
|
|
; VI-NEXT: s_lshr_b32 s76, s26, 16
|
|
; VI-NEXT: s_lshr_b32 s77, s40, 16
|
|
; VI-NEXT: s_lshr_b32 s78, s41, 16
|
|
; VI-NEXT: s_lshr_b32 s79, s42, 16
|
|
; VI-NEXT: .LBB45_3: ; %end
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s42
|
|
; VI-NEXT: s_lshl_b32 s5, s79, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s41
|
|
; VI-NEXT: s_lshl_b32 s41, s78, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s41
|
|
; VI-NEXT: s_and_b32 s40, 0xffff, s40
|
|
; VI-NEXT: s_lshl_b32 s41, s77, 16
|
|
; VI-NEXT: s_or_b32 s40, s40, s41
|
|
; VI-NEXT: s_and_b32 s26, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s41, s76, 16
|
|
; VI-NEXT: s_or_b32 s26, s26, s41
|
|
; VI-NEXT: s_and_b32 s25, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s41, s75, 16
|
|
; VI-NEXT: s_or_b32 s25, s25, s41
|
|
; VI-NEXT: s_and_b32 s24, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s41, s74, 16
|
|
; VI-NEXT: s_or_b32 s24, s24, s41
|
|
; VI-NEXT: s_and_b32 s23, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s41, s73, 16
|
|
; VI-NEXT: s_or_b32 s23, s23, s41
|
|
; VI-NEXT: s_and_b32 s22, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s41, s72, 16
|
|
; VI-NEXT: s_or_b32 s22, s22, s41
|
|
; VI-NEXT: s_and_b32 s21, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s41, s63, 16
|
|
; VI-NEXT: s_or_b32 s21, s21, s41
|
|
; VI-NEXT: s_and_b32 s20, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s41, s62, 16
|
|
; VI-NEXT: s_or_b32 s20, s20, s41
|
|
; VI-NEXT: s_and_b32 s19, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s41, s61, 16
|
|
; VI-NEXT: s_or_b32 s19, s19, s41
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s41, s60, 16
|
|
; VI-NEXT: s_or_b32 s18, s18, s41
|
|
; VI-NEXT: s_and_b32 s17, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s41, s59, 16
|
|
; VI-NEXT: s_or_b32 s17, s17, s41
|
|
; VI-NEXT: s_and_b32 s16, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s41, s58, 16
|
|
; VI-NEXT: s_or_b32 s16, s16, s41
|
|
; VI-NEXT: s_and_b32 s15, 0xffff, s15
|
|
; VI-NEXT: s_lshl_b32 s41, s57, 16
|
|
; VI-NEXT: s_or_b32 s15, s15, s41
|
|
; VI-NEXT: s_and_b32 s14, 0xffff, s14
|
|
; VI-NEXT: s_lshl_b32 s41, s56, 16
|
|
; VI-NEXT: s_or_b32 s14, s14, s41
|
|
; VI-NEXT: s_and_b32 s13, 0xffff, s13
|
|
; VI-NEXT: s_lshl_b32 s41, s47, 16
|
|
; VI-NEXT: s_or_b32 s13, s13, s41
|
|
; VI-NEXT: s_and_b32 s12, 0xffff, s12
|
|
; VI-NEXT: s_lshl_b32 s41, s46, 16
|
|
; VI-NEXT: s_or_b32 s12, s12, s41
|
|
; VI-NEXT: s_and_b32 s11, 0xffff, s11
|
|
; VI-NEXT: s_lshl_b32 s41, s45, 16
|
|
; VI-NEXT: s_or_b32 s11, s11, s41
|
|
; VI-NEXT: s_and_b32 s10, 0xffff, s10
|
|
; VI-NEXT: s_lshl_b32 s41, s44, 16
|
|
; VI-NEXT: s_or_b32 s10, s10, s41
|
|
; VI-NEXT: s_and_b32 s9, 0xffff, s9
|
|
; VI-NEXT: s_lshl_b32 s41, s43, 16
|
|
; VI-NEXT: s_and_b32 s8, 0xffff, s8
|
|
; VI-NEXT: s_lshl_b32 s29, s29, 16
|
|
; VI-NEXT: s_and_b32 s6, 0xffff, s6
|
|
; VI-NEXT: s_lshl_b32 s28, s28, 16
|
|
; VI-NEXT: s_and_b32 s7, 0xffff, s7
|
|
; VI-NEXT: s_lshl_b32 s27, s27, 16
|
|
; VI-NEXT: s_or_b32 s9, s9, s41
|
|
; VI-NEXT: s_or_b32 s8, s8, s29
|
|
; VI-NEXT: s_or_b32 s6, s6, s28
|
|
; VI-NEXT: s_or_b32 s7, s7, s27
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s40
|
|
; VI-NEXT: v_mov_b32_e32 v3, s26
|
|
; VI-NEXT: v_mov_b32_e32 v4, s25
|
|
; VI-NEXT: v_mov_b32_e32 v5, s24
|
|
; VI-NEXT: v_mov_b32_e32 v6, s23
|
|
; VI-NEXT: v_mov_b32_e32 v7, s22
|
|
; VI-NEXT: v_mov_b32_e32 v8, s21
|
|
; VI-NEXT: v_mov_b32_e32 v9, s20
|
|
; VI-NEXT: v_mov_b32_e32 v10, s19
|
|
; VI-NEXT: v_mov_b32_e32 v11, s18
|
|
; VI-NEXT: v_mov_b32_e32 v12, s17
|
|
; VI-NEXT: v_mov_b32_e32 v13, s16
|
|
; VI-NEXT: v_mov_b32_e32 v14, s15
|
|
; VI-NEXT: v_mov_b32_e32 v15, s14
|
|
; VI-NEXT: v_mov_b32_e32 v16, s13
|
|
; VI-NEXT: v_mov_b32_e32 v17, s12
|
|
; VI-NEXT: v_mov_b32_e32 v18, s11
|
|
; VI-NEXT: v_mov_b32_e32 v19, s10
|
|
; VI-NEXT: v_mov_b32_e32 v20, s9
|
|
; VI-NEXT: v_mov_b32_e32 v21, s8
|
|
; VI-NEXT: v_mov_b32_e32 v22, s6
|
|
; VI-NEXT: v_mov_b32_e32 v23, s7
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB45_4:
|
|
; 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: ; implicit-def: $sgpr43
|
|
; VI-NEXT: ; implicit-def: $sgpr29
|
|
; VI-NEXT: ; implicit-def: $sgpr28
|
|
; VI-NEXT: ; implicit-def: $sgpr27
|
|
; VI-NEXT: s_branch .LBB45_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v12i64_to_v48f16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s24
|
|
; GFX9-NEXT: v_readfirstlane_b32 s6, v11
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s25
|
|
; GFX9-NEXT: v_readfirstlane_b32 s7, v12
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s26
|
|
; GFX9-NEXT: v_readfirstlane_b32 s8, v13
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s27
|
|
; GFX9-NEXT: v_readfirstlane_b32 s9, v14
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s28
|
|
; GFX9-NEXT: v_readfirstlane_b32 s10, v15
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s29
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_readfirstlane_b32 s11, v16
|
|
; GFX9-NEXT: v_readfirstlane_b32 s12, v17
|
|
; GFX9-NEXT: v_readfirstlane_b32 s13, v18
|
|
; GFX9-NEXT: v_readfirstlane_b32 s14, v19
|
|
; GFX9-NEXT: v_readfirstlane_b32 s15, v11
|
|
; GFX9-NEXT: v_readfirstlane_b32 s16, v12
|
|
; GFX9-NEXT: v_readfirstlane_b32 s17, v13
|
|
; GFX9-NEXT: v_readfirstlane_b32 s18, v14
|
|
; GFX9-NEXT: v_readfirstlane_b32 s19, v15
|
|
; GFX9-NEXT: v_readfirstlane_b32 s20, v0
|
|
; GFX9-NEXT: v_readfirstlane_b32 s21, v1
|
|
; GFX9-NEXT: v_readfirstlane_b32 s22, v2
|
|
; GFX9-NEXT: v_readfirstlane_b32 s23, v3
|
|
; GFX9-NEXT: v_readfirstlane_b32 s24, v4
|
|
; GFX9-NEXT: v_readfirstlane_b32 s25, v5
|
|
; GFX9-NEXT: v_readfirstlane_b32 s26, v6
|
|
; GFX9-NEXT: v_readfirstlane_b32 s27, v7
|
|
; GFX9-NEXT: v_readfirstlane_b32 s28, v8
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_readfirstlane_b32 s29, v9
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB45_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s44, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s6, 16
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB45_3
|
|
; GFX9-NEXT: .LBB45_2: ; %cmp.true
|
|
; 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_add_u32 s14, s14, 3
|
|
; GFX9-NEXT: s_addc_u32 s15, s15, 0
|
|
; GFX9-NEXT: s_add_u32 s12, s12, 3
|
|
; GFX9-NEXT: s_addc_u32 s13, s13, 0
|
|
; GFX9-NEXT: s_add_u32 s10, s10, 3
|
|
; GFX9-NEXT: s_addc_u32 s11, s11, 0
|
|
; GFX9-NEXT: s_add_u32 s8, s8, 3
|
|
; GFX9-NEXT: s_addc_u32 s9, s9, 0
|
|
; GFX9-NEXT: s_add_u32 s6, s6, 3
|
|
; GFX9-NEXT: s_addc_u32 s7, s7, 0
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX9-NEXT: s_lshr_b32 s44, s25, 16
|
|
; GFX9-NEXT: s_lshr_b32 s45, s24, 16
|
|
; GFX9-NEXT: s_lshr_b32 s46, s23, 16
|
|
; GFX9-NEXT: s_lshr_b32 s47, s22, 16
|
|
; GFX9-NEXT: s_lshr_b32 s56, s21, 16
|
|
; GFX9-NEXT: s_lshr_b32 s57, s20, 16
|
|
; GFX9-NEXT: s_lshr_b32 s58, s19, 16
|
|
; GFX9-NEXT: s_lshr_b32 s59, s18, 16
|
|
; GFX9-NEXT: s_lshr_b32 s60, s17, 16
|
|
; GFX9-NEXT: s_lshr_b32 s61, s16, 16
|
|
; GFX9-NEXT: s_lshr_b32 s62, s15, 16
|
|
; GFX9-NEXT: s_lshr_b32 s63, s14, 16
|
|
; GFX9-NEXT: s_lshr_b32 s72, s13, 16
|
|
; GFX9-NEXT: s_lshr_b32 s73, s12, 16
|
|
; GFX9-NEXT: s_lshr_b32 s74, s11, 16
|
|
; GFX9-NEXT: s_lshr_b32 s75, s10, 16
|
|
; GFX9-NEXT: s_lshr_b32 s76, s9, 16
|
|
; GFX9-NEXT: s_lshr_b32 s77, s8, 16
|
|
; GFX9-NEXT: s_lshr_b32 s78, s7, 16
|
|
; GFX9-NEXT: s_lshr_b32 s79, s6, 16
|
|
; GFX9-NEXT: .LBB45_3: ; %end
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s6, s79
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s5, s7, s78
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s8, s77
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s9, s76
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s10, s75
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s11, s74
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s12, s73
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s13, s72
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s14, s63
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s15, s62
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s16, s61
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s17, s60
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s59
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s58
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s57
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s56
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s47
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s46
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s45
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s44
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s4
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s5
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s27
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB45_4:
|
|
; 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: ; implicit-def: $sgpr43
|
|
; GFX9-NEXT: ; implicit-def: $sgpr42
|
|
; GFX9-NEXT: ; implicit-def: $sgpr41
|
|
; GFX9-NEXT: ; implicit-def: $sgpr40
|
|
; GFX9-NEXT: s_branch .LBB45_2
|
|
;
|
|
; GFX11-LABEL: bitcast_v12i64_to_v48f16_scalar:
|
|
; GFX11: ; %bb.0:
|
|
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-NEXT: v_dual_mov_b32 v7, s0 :: v_dual_mov_b32 v8, s1
|
|
; GFX11-NEXT: v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
|
|
; GFX11-NEXT: v_dual_mov_b32 v11, s16 :: v_dual_mov_b32 v12, s17
|
|
; GFX11-NEXT: v_dual_mov_b32 v13, s18 :: v_dual_mov_b32 v14, s19
|
|
; GFX11-NEXT: v_dual_mov_b32 v15, s20 :: v_dual_mov_b32 v16, s21
|
|
; GFX11-NEXT: v_dual_mov_b32 v17, s22 :: v_dual_mov_b32 v18, s23
|
|
; GFX11-NEXT: v_dual_mov_b32 v19, s24 :: v_dual_mov_b32 v20, s25
|
|
; GFX11-NEXT: v_dual_mov_b32 v21, s26 :: v_dual_mov_b32 v22, s27
|
|
; GFX11-NEXT: v_dual_mov_b32 v23, s28 :: v_dual_mov_b32 v24, s29
|
|
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-NEXT: v_readfirstlane_b32 s0, v7
|
|
; GFX11-NEXT: v_readfirstlane_b32 s1, v8
|
|
; GFX11-NEXT: v_readfirstlane_b32 s2, v9
|
|
; GFX11-NEXT: v_readfirstlane_b32 s3, v10
|
|
; GFX11-NEXT: v_readfirstlane_b32 s4, v11
|
|
; GFX11-NEXT: v_readfirstlane_b32 s5, v12
|
|
; GFX11-NEXT: v_readfirstlane_b32 s6, v13
|
|
; GFX11-NEXT: v_readfirstlane_b32 s7, v14
|
|
; GFX11-NEXT: v_readfirstlane_b32 s8, v15
|
|
; GFX11-NEXT: v_readfirstlane_b32 s9, v16
|
|
; GFX11-NEXT: v_readfirstlane_b32 s10, v17
|
|
; GFX11-NEXT: v_readfirstlane_b32 s11, v18
|
|
; GFX11-NEXT: v_readfirstlane_b32 s12, v19
|
|
; GFX11-NEXT: v_readfirstlane_b32 s13, v20
|
|
; GFX11-NEXT: v_readfirstlane_b32 s14, v21
|
|
; GFX11-NEXT: v_readfirstlane_b32 s15, v22
|
|
; GFX11-NEXT: v_readfirstlane_b32 s16, v23
|
|
; GFX11-NEXT: v_readfirstlane_b32 s17, v24
|
|
; GFX11-NEXT: v_readfirstlane_b32 s18, v0
|
|
; GFX11-NEXT: v_readfirstlane_b32 s19, v1
|
|
; GFX11-NEXT: v_readfirstlane_b32 s20, v2
|
|
; GFX11-NEXT: v_readfirstlane_b32 s21, v3
|
|
; GFX11-NEXT: v_readfirstlane_b32 s23, v4
|
|
; GFX11-NEXT: v_readfirstlane_b32 s22, v5
|
|
; GFX11-NEXT: s_mov_b32 s74, 0
|
|
; GFX11-NEXT: s_and_b32 s24, vcc_lo, exec_lo
|
|
; GFX11-NEXT: s_cbranch_scc0 .LBB45_4
|
|
; GFX11-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-NEXT: s_lshr_b32 s24, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s25, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s26, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s27, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s28, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s29, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s15, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s14, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, s0, 16
|
|
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s74
|
|
; GFX11-NEXT: s_cbranch_vccnz .LBB45_3
|
|
; GFX11-NEXT: .LBB45_2: ; %cmp.true
|
|
; GFX11-NEXT: s_add_u32 s23, s23, 3
|
|
; GFX11-NEXT: s_addc_u32 s22, s22, 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 s14, s14, 3
|
|
; GFX11-NEXT: s_addc_u32 s15, s15, 0
|
|
; GFX11-NEXT: s_add_u32 s12, s12, 3
|
|
; GFX11-NEXT: s_addc_u32 s13, s13, 0
|
|
; GFX11-NEXT: s_add_u32 s10, s10, 3
|
|
; GFX11-NEXT: s_addc_u32 s11, s11, 0
|
|
; GFX11-NEXT: s_add_u32 s8, s8, 3
|
|
; GFX11-NEXT: s_addc_u32 s9, s9, 0
|
|
; GFX11-NEXT: s_add_u32 s6, s6, 3
|
|
; GFX11-NEXT: s_addc_u32 s7, s7, 0
|
|
; GFX11-NEXT: s_add_u32 s4, s4, 3
|
|
; GFX11-NEXT: s_addc_u32 s5, s5, 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 s24, s22, 16
|
|
; GFX11-NEXT: s_lshr_b32 s25, s23, 16
|
|
; GFX11-NEXT: s_lshr_b32 s26, s21, 16
|
|
; GFX11-NEXT: s_lshr_b32 s27, s20, 16
|
|
; GFX11-NEXT: s_lshr_b32 s28, s19, 16
|
|
; GFX11-NEXT: s_lshr_b32 s29, s18, 16
|
|
; GFX11-NEXT: s_lshr_b32 s40, s17, 16
|
|
; GFX11-NEXT: s_lshr_b32 s41, s16, 16
|
|
; GFX11-NEXT: s_lshr_b32 s42, s15, 16
|
|
; GFX11-NEXT: s_lshr_b32 s43, s14, 16
|
|
; GFX11-NEXT: s_lshr_b32 s44, s13, 16
|
|
; GFX11-NEXT: s_lshr_b32 s45, s12, 16
|
|
; GFX11-NEXT: s_lshr_b32 s46, s11, 16
|
|
; GFX11-NEXT: s_lshr_b32 s47, s10, 16
|
|
; GFX11-NEXT: s_lshr_b32 s56, s9, 16
|
|
; GFX11-NEXT: s_lshr_b32 s57, s8, 16
|
|
; GFX11-NEXT: s_lshr_b32 s58, s7, 16
|
|
; GFX11-NEXT: s_lshr_b32 s59, s6, 16
|
|
; GFX11-NEXT: s_lshr_b32 s60, s5, 16
|
|
; GFX11-NEXT: s_lshr_b32 s61, s4, 16
|
|
; GFX11-NEXT: s_lshr_b32 s62, s3, 16
|
|
; GFX11-NEXT: s_lshr_b32 s63, s2, 16
|
|
; GFX11-NEXT: s_lshr_b32 s72, s1, 16
|
|
; GFX11-NEXT: s_lshr_b32 s73, 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, s73
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s72
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s63
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s62
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s61
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s60
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s59
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s58
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s57
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s56
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s47
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s46
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s45
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s44
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s43
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s42
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s41
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s40
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s29
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s28
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s27
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s26
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s25
|
|
; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s24
|
|
; 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, s23 :: v_dual_mov_b32 v23, s22
|
|
; GFX11-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-NEXT: .LBB45_4:
|
|
; 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: $sgpr29
|
|
; GFX11-NEXT: ; implicit-def: $sgpr28
|
|
; GFX11-NEXT: ; implicit-def: $sgpr27
|
|
; GFX11-NEXT: ; implicit-def: $sgpr26
|
|
; GFX11-NEXT: ; implicit-def: $sgpr25
|
|
; GFX11-NEXT: ; implicit-def: $sgpr24
|
|
; 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 <12 x i64> %a, splat (i64 3)
|
|
%a2 = bitcast <12 x i64> %a1 to <48 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x i64> %a to <48 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x half> %phi
|
|
}
|
|
|
|
define <12 x i64> @bitcast_v48f16_to_v12i64(<48 x half> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v48f16_to_v12i64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v14
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:24
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:20
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:32
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:28
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:48
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:44
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:56
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:52
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:64
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:60
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v1
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, v5
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v6
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v11
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, v12
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v15
|
|
; 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_cvt_f16_f32_e32 v0, v20
|
|
; 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_cvt_f16_f32_e32 v0, v23
|
|
; 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_cvt_f16_f32_e32 v0, v22
|
|
; 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_cvt_f16_f32_e32 v0, v25
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v24
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v27
|
|
; 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_cvt_f16_f32_e32 v0, v26
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v29
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v28
|
|
; 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_cvt_f16_f32_e32 v0, v30
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v31
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v40, v63
|
|
; 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_cvt_f16_f32_e32 v0, v41
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v62
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v42
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, v61
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v43
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v60
|
|
; 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_cvt_f16_f32_e32 v0, v44
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v59
|
|
; 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_cvt_f16_f32_e32 v0, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v45, v58
|
|
; 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_cvt_f16_f32_e32 v0, v46
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, v57
|
|
; 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_cvt_f16_f32_e32 v0, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v56
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 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 v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v33
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v55
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v49
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v45
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v43
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v41
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_or_b32_e32 v0, v54, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v52, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v50, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v48, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v38, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v36, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v34, v6
|
|
; SI-NEXT: v_or_b32_e32 v21, v44, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v42, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v40, v23
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; 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: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; 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: buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v20, v46, v20
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; 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: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v55
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v53
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v52
|
|
; 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, v51
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v48
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v38
|
|
; 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, v49
|
|
; 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, v36
|
|
; 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, v34
|
|
; 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, v39
|
|
; 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 v21, v47
|
|
; 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 v22, v46
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; 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, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v44
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v41
|
|
; 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, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v40
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; 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, v33
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v12
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v11, v13, v11
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; 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: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; 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, v14, v12
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; 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: 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_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v15
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_or_b32_e32 v14, v16, v14
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v18
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_or_b32_e32 v17, v19, v17
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v45
|
|
; SI-NEXT: v_or_b32_e32 v20, v22, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v43
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v23, v25, v23
|
|
; 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 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v48f16_to_v12i64:
|
|
; 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: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_mov_b32_e32 v32, v23
|
|
; VI-NEXT: v_mov_b32_e32 v33, v22
|
|
; VI-NEXT: v_mov_b32_e32 v34, v21
|
|
; VI-NEXT: v_mov_b32_e32 v35, v20
|
|
; VI-NEXT: v_mov_b32_e32 v36, v19
|
|
; VI-NEXT: v_mov_b32_e32 v37, v18
|
|
; VI-NEXT: v_mov_b32_e32 v38, v17
|
|
; VI-NEXT: v_mov_b32_e32 v39, v16
|
|
; VI-NEXT: v_mov_b32_e32 v48, v15
|
|
; VI-NEXT: v_mov_b32_e32 v49, v14
|
|
; VI-NEXT: v_mov_b32_e32 v50, v13
|
|
; VI-NEXT: v_mov_b32_e32 v51, v12
|
|
; VI-NEXT: v_mov_b32_e32 v52, v11
|
|
; VI-NEXT: v_mov_b32_e32 v53, v10
|
|
; VI-NEXT: v_mov_b32_e32 v54, v9
|
|
; VI-NEXT: v_mov_b32_e32 v55, v8
|
|
; VI-NEXT: v_mov_b32_e32 v40, v7
|
|
; VI-NEXT: v_mov_b32_e32 v41, v6
|
|
; VI-NEXT: v_mov_b32_e32 v42, v5
|
|
; VI-NEXT: v_mov_b32_e32 v43, v4
|
|
; VI-NEXT: v_mov_b32_e32 v44, v3
|
|
; VI-NEXT: v_mov_b32_e32 v45, v2
|
|
; VI-NEXT: v_mov_b32_e32 v46, v1
|
|
; VI-NEXT: v_mov_b32_e32 v47, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 v23, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v23, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v23, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v23, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v23, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v23, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v23, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v23, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v23, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v23, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v23, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v23, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v23, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v23, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v23, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v23, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v23, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v23, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v23, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v23, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v23, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v23, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v23, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v23, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v44, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v43, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v42, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v41, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v40, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v55, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v53, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v52, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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 v23, 0x200
|
|
; VI-NEXT: v_add_f16_sdwa v0, v47, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v47
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_add_f16_sdwa v1, v46, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v2, 0x200, v46
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; VI-NEXT: v_add_f16_sdwa v2, v45, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v45
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v3, v44, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v44
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_add_f16_sdwa v4, v43, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v43
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_add_f16_sdwa v5, v42, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v42
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_add_f16_sdwa v6, v41, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v41
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_add_f16_sdwa v7, v40, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v40
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_add_f16_sdwa v8, v55, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v55
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_add_f16_sdwa v9, v54, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, 0x200, v54
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, 0x200, v53
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_add_f16_sdwa v11, v52, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, 0x200, v52
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_add_f16_sdwa v12, v51, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, 0x200, v51
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_add_f16_sdwa v13, v50, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v50
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_add_f16_sdwa v14, v49, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v48, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v39, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v38, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v37, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v36, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v35, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v34, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB46_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48f16_to_v12i64:
|
|
; 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 v33, v22
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v13
|
|
; 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, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v12
|
|
; 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, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v11
|
|
; 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, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v10
|
|
; 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, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v1
|
|
; 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, v53
|
|
; 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, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v47
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 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:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; 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: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; 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 v9, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: s_movk_i32 s7, 0x200
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, 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(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, 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: .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_v48f16_to_v12i64:
|
|
; 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, v24
|
|
; 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: .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_v48f16_to_v12i64:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v24
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v53, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v55, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v64, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v52, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v51, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v50, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v49, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v48, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v39, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v38, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v37, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v36, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v35, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v34, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v33, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v32, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v31, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v30, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v29, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v28, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v27, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v26, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v25, v23, 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: .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 <48 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <48 x half> %a1 to <12 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x half> %a to <12 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x i64> %phi
|
|
}
|
|
|
|
define inreg <12 x i64> @bitcast_v48f16_to_v12i64_scalar(<48 x half> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v48f16_to_v12i64_scalar:
|
|
; 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:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v62, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v63, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v61, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v60, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v59, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v58, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v57, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v56, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v45, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, s17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, s16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, s19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, s18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, s21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, s20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, s23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, s22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, s25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, s24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, s27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v30, s28
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v31
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v34
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v35
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, s26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, s29
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v40
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: s_cbranch_scc0 .LBB47_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v0, v8, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v7, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v6, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v5, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v34
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v62
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v52
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v48
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v60
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v58
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v46
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v4, v54, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v35, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v30, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v55, v7
|
|
; SI-NEXT: v_or_b32_e32 v8, v63, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v43, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v49, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v39, v11
|
|
; SI-NEXT: v_or_b32_e32 v12, v37, v12
|
|
; SI-NEXT: v_or_b32_e32 v13, v61, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v33, v14
|
|
; SI-NEXT: v_or_b32_e32 v15, v59, v15
|
|
; SI-NEXT: v_or_b32_e32 v16, v57, v16
|
|
; SI-NEXT: v_or_b32_e32 v17, v47, v17
|
|
; SI-NEXT: v_or_b32_e32 v18, v45, v18
|
|
; SI-NEXT: v_or_b32_e32 v19, v25, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v41, v20
|
|
; SI-NEXT: v_or_b32_e32 v21, v28, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v26, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v31, v23
|
|
; SI-NEXT: s_cbranch_execnz .LBB47_3
|
|
; SI-NEXT: .LBB47_2: ; %cmp.true
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v30
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v55
|
|
; 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_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; 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 v9, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v49
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v48
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v60
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v33
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; 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_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v59
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v56
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v47
|
|
; 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, v45
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v42
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v41
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v28
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, 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 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: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_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 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; 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, v51
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; 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, v34
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; 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, v62
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; 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, v53
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; 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, v52
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; 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, v50
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v38
|
|
; SI-NEXT: v_or_b32_e32 v11, v13, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v36
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v61
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v32
|
|
; SI-NEXT: v_or_b32_e32 v14, v16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v58
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v57
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v46
|
|
; SI-NEXT: v_or_b32_e32 v17, v19, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v44
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v31
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v29
|
|
; SI-NEXT: v_or_b32_e32 v20, v22, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v27
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v23, v25, v23
|
|
; SI-NEXT: .LBB47_3: ; %end
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB47_4:
|
|
; SI-NEXT: v_mov_b32_e32 v40, v31
|
|
; SI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v54, v33
|
|
; SI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v51, v32
|
|
; SI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v34, v59
|
|
; SI-NEXT: v_mov_b32_e32 v59, v46
|
|
; SI-NEXT: v_mov_b32_e32 v46, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v24
|
|
; SI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v35, v60
|
|
; SI-NEXT: v_mov_b32_e32 v60, v47
|
|
; SI-NEXT: v_mov_b32_e32 v47, v42
|
|
; SI-NEXT: v_mov_b32_e32 v42, v26
|
|
; SI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v55, v52
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_mov_b32_e32 v53, v43
|
|
; SI-NEXT: v_mov_b32_e32 v52, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v61
|
|
; SI-NEXT: v_mov_b32_e32 v61, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v25
|
|
; SI-NEXT: v_mov_b32_e32 v43, v27
|
|
; SI-NEXT: v_mov_b32_e32 v33, v62
|
|
; SI-NEXT: v_mov_b32_e32 v62, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v44
|
|
; SI-NEXT: v_mov_b32_e32 v44, v28
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v63, v58
|
|
; SI-NEXT: v_mov_b32_e32 v58, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v29
|
|
; SI-NEXT: v_mov_b32_e32 v32, v30
|
|
; 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: v_mov_b32_e32 v29, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v58
|
|
; SI-NEXT: v_mov_b32_e32 v58, v63
|
|
; SI-NEXT: v_mov_b32_e32 v27, v43
|
|
; SI-NEXT: v_mov_b32_e32 v25, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v61
|
|
; SI-NEXT: v_mov_b32_e32 v61, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v52
|
|
; SI-NEXT: v_mov_b32_e32 v43, v53
|
|
; SI-NEXT: v_mov_b32_e32 v52, v55
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v30, v32
|
|
; SI-NEXT: v_mov_b32_e32 v28, v44
|
|
; SI-NEXT: v_mov_b32_e32 v44, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v62
|
|
; SI-NEXT: v_mov_b32_e32 v62, v33
|
|
; SI-NEXT: v_mov_b32_e32 v26, v42
|
|
; SI-NEXT: v_mov_b32_e32 v42, v47
|
|
; SI-NEXT: v_mov_b32_e32 v47, v60
|
|
; SI-NEXT: v_mov_b32_e32 v60, v35
|
|
; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v24, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v46
|
|
; SI-NEXT: v_mov_b32_e32 v46, v59
|
|
; SI-NEXT: v_mov_b32_e32 v59, v34
|
|
; SI-NEXT: v_mov_b32_e32 v32, v51
|
|
; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v33, v54
|
|
; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v31, v40
|
|
; SI-NEXT: s_branch .LBB47_2
|
|
;
|
|
; VI-LABEL: bitcast_v48f16_to_v12i64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v32, v9
|
|
; VI-NEXT: v_mov_b32_e32 v33, v8
|
|
; VI-NEXT: v_mov_b32_e32 v34, v7
|
|
; VI-NEXT: v_mov_b32_e32 v35, v6
|
|
; VI-NEXT: v_mov_b32_e32 v36, v5
|
|
; VI-NEXT: v_mov_b32_e32 v37, v4
|
|
; VI-NEXT: v_mov_b32_e32 v38, v3
|
|
; VI-NEXT: v_mov_b32_e32 v39, v2
|
|
; VI-NEXT: v_mov_b32_e32 v48, v1
|
|
; VI-NEXT: v_mov_b32_e32 v49, v0
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: s_cbranch_scc0 .LBB47_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s43, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s44, s42, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s44
|
|
; VI-NEXT: s_and_b32 s44, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s45, s41, 16
|
|
; VI-NEXT: s_or_b32 s44, s44, s45
|
|
; VI-NEXT: s_and_b32 s45, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s46, s40, 16
|
|
; VI-NEXT: s_or_b32 s45, s45, s46
|
|
; VI-NEXT: s_and_b32 s46, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s47, s15, 16
|
|
; VI-NEXT: s_or_b32 s46, s46, s47
|
|
; VI-NEXT: s_and_b32 s47, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s56, s14, 16
|
|
; VI-NEXT: v_mov_b32_e32 v0, 16
|
|
; VI-NEXT: s_or_b32 s47, s47, s56
|
|
; VI-NEXT: s_and_b32 s56, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s57, s13, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s56, s56, s57
|
|
; VI-NEXT: s_and_b32 s57, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s58, s12, 16
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s57, s57, s58
|
|
; VI-NEXT: s_and_b32 s58, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s59, s11, 16
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s58, s58, s59
|
|
; VI-NEXT: s_and_b32 s59, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s60, s10, 16
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s59, s59, s60
|
|
; VI-NEXT: s_and_b32 s60, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s61, s9, 16
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s60, s60, s61
|
|
; VI-NEXT: s_and_b32 s61, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s62, s8, 16
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s61, s61, s62
|
|
; VI-NEXT: s_and_b32 s62, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s63, s7, 16
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s62, s62, s63
|
|
; VI-NEXT: s_and_b32 s63, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s72, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s63, s63, s72
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v0, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s44
|
|
; VI-NEXT: v_mov_b32_e32 v3, s45
|
|
; VI-NEXT: v_mov_b32_e32 v4, s46
|
|
; VI-NEXT: v_mov_b32_e32 v5, s47
|
|
; VI-NEXT: v_mov_b32_e32 v6, s56
|
|
; VI-NEXT: v_mov_b32_e32 v7, s57
|
|
; VI-NEXT: v_mov_b32_e32 v8, s58
|
|
; VI-NEXT: v_mov_b32_e32 v9, s59
|
|
; VI-NEXT: v_mov_b32_e32 v10, s60
|
|
; VI-NEXT: v_mov_b32_e32 v11, s61
|
|
; VI-NEXT: v_mov_b32_e32 v12, s62
|
|
; VI-NEXT: v_mov_b32_e32 v13, s63
|
|
; VI-NEXT: s_cbranch_execnz .LBB47_3
|
|
; VI-NEXT: .LBB47_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v13, 0x200
|
|
; VI-NEXT: v_mov_b32_e32 v0, s43
|
|
; VI-NEXT: v_add_f16_sdwa v0, v0, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, s16, v13
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_mov_b32_e32 v1, s42
|
|
; VI-NEXT: v_add_f16_sdwa v1, v1, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v2, s17, v13
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; VI-NEXT: v_mov_b32_e32 v2, s41
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s18, v13
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v3, s40
|
|
; VI-NEXT: v_add_f16_sdwa v3, v3, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, s19, v13
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_mov_b32_e32 v4, s15
|
|
; VI-NEXT: v_add_f16_sdwa v4, v4, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, s20, v13
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_mov_b32_e32 v5, s14
|
|
; VI-NEXT: v_add_f16_sdwa v5, v5, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, s21, v13
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_mov_b32_e32 v6, s13
|
|
; VI-NEXT: v_add_f16_sdwa v6, v6, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, s22, v13
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_mov_b32_e32 v7, s12
|
|
; VI-NEXT: v_add_f16_sdwa v7, v7, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, s23, v13
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_mov_b32_e32 v8, s11
|
|
; VI-NEXT: v_add_f16_sdwa v8, v8, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, s24, v13
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_mov_b32_e32 v9, s10
|
|
; VI-NEXT: v_add_f16_sdwa v9, v9, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, s25, v13
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_mov_b32_e32 v10, s9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v10, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, s26, v13
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_mov_b32_e32 v11, s8
|
|
; VI-NEXT: v_add_f16_sdwa v11, v11, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, s27, v13
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_mov_b32_e32 v12, s7
|
|
; VI-NEXT: v_add_f16_sdwa v12, v12, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, s28, v13
|
|
; VI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; VI-NEXT: v_mov_b32_e32 v14, s6
|
|
; VI-NEXT: v_add_f16_sdwa v14, v14, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, s29, v13
|
|
; VI-NEXT: v_mov_b32_e32 v23, 0x200
|
|
; VI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; VI-NEXT: v_add_f16_sdwa v14, v49, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v48, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v39, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v38, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v37, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v36, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v35, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v34, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB47_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB47_4:
|
|
; 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_branch .LBB47_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v48f16_to_v12i64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v0
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, 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: buffer_store_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v32
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v33
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v49
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s17, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s18, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s19, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s20, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s21, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s22, s12
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s23, s13
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s24, s14
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s25, s15
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s29, s40
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB47_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v33
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s19
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB47_3
|
|
; GFX9-NEXT: .LBB47_2: ; %cmp.true
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v49
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v48
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v39
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v38
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v37
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v36
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v35
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v34
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v33
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v32
|
|
; GFX9-NEXT: s_movk_i32 s4, 0x200
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, 0x200
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v23
|
|
; GFX9-NEXT: v_pk_add_f16 v0, s6, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, s7, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, s8, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, s9, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, s10, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, s11, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, s12, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, s13, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, s14, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v9, s15, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, s16, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, s17, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, s18, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, s19, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, v14, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, v15, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, v16, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, v17, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, v18, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, v19, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, v20, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, v21, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, v22, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, v23, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB47_3: ; %end
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB47_4:
|
|
; 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_branch .LBB47_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v12i64_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB47_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; 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, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB47_3
|
|
; GFX11-TRUE16-NEXT: .LBB47_2: ; %cmp.true
|
|
; 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, s4 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, s17 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: .LBB47_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB47_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB47_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v48f16_to_v12i64_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s4, s0, s4
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s44
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s16, s5
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s17, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s18, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s19, s8
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s20, s9
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s21, s10
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s22, s11
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s23, s12
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s24, s13
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s25, s14
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s26, s43
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB47_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB47_3
|
|
; GFX11-FAKE16-NEXT: .LBB47_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, s4 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 v2, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, s17 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: .LBB47_3: ; %end
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB47_4:
|
|
; GFX11-FAKE16-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
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB47_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <48 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <48 x half> %a1 to <12 x i64>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x half> %a to <12 x i64>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x i64> %phi
|
|
}
|
|
|
|
define <48 x i16> @bitcast_v12f64_to_v48i16(<12 x double> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v12f64_to_v48i16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v25
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; 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 v25, v24, v23, 16
|
|
; SI-NEXT: v_alignbit_b32 v26, v22, v21, 16
|
|
; SI-NEXT: v_alignbit_b32 v27, v20, v19, 16
|
|
; SI-NEXT: v_alignbit_b32 v28, v18, v17, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v16, v15, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v14, v13, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v12, v11, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v10, v9, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v8, v7, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v6, v5, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v4, v3, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v2, v1, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v2
|
|
; 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[1:2], v[1:2], 1.0
|
|
; SI-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
|
|
; SI-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
|
|
; SI-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
|
|
; SI-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
|
|
; SI-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
|
|
; SI-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
|
|
; SI-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
|
|
; SI-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
|
|
; SI-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
|
|
; SI-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
|
|
; SI-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
|
|
; SI-NEXT: v_alignbit_b32 v25, v24, v23, 16
|
|
; SI-NEXT: v_alignbit_b32 v26, v22, v21, 16
|
|
; SI-NEXT: v_alignbit_b32 v27, v20, v19, 16
|
|
; SI-NEXT: v_alignbit_b32 v28, v18, v17, 16
|
|
; SI-NEXT: v_alignbit_b32 v29, v16, v15, 16
|
|
; SI-NEXT: v_alignbit_b32 v31, v14, v13, 16
|
|
; SI-NEXT: v_alignbit_b32 v33, v12, v11, 16
|
|
; SI-NEXT: v_alignbit_b32 v35, v10, v9, 16
|
|
; SI-NEXT: v_alignbit_b32 v38, v8, v7, 16
|
|
; SI-NEXT: v_alignbit_b32 v48, v6, v5, 16
|
|
; SI-NEXT: v_alignbit_b32 v50, v4, v3, 16
|
|
; SI-NEXT: v_alignbit_b32 v53, v2, v1, 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v2
|
|
; SI-NEXT: .LBB48_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v53, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v53
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 4, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 8, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 12, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 16, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 20, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 24, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 28, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 32, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 36, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 40, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v31
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 52, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v29
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 56, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v37
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 60, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 64, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v27
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x4c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x50, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x54, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x58, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v30
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v12f64_to_v48i16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 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[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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; VI-NEXT: .LBB48_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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_lshlrev_b32_e32 v27, 16, v27
|
|
; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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_or_b32_sdwa v20, v20, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v12f64_to_v48i16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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[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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v54, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v53, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v52, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v51, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v50, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v49, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v48, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v39, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v37, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v36, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v35, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v34, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v33, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v32, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v31, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v30, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v29, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v28, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v27, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v26, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v25, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s4
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v12f64_to_v48i16:
|
|
; 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, v24
|
|
; 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[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_v12f64_to_v48i16:
|
|
; 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, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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[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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v53, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v52, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v51, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v50, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v49, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v48, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v39, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v37, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v35, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v33, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v32, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v31, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v30, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v29, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v28, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v27, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v26, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v25, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v24, v23, 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 <12 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <12 x double> %a1 to <48 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x double> %a to <48 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x i16> %phi
|
|
}
|
|
|
|
define inreg <48 x i16> @bitcast_v12f64_to_v48i16_scalar(<12 x double> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v12f64_to_v48i16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
|
|
; SI-NEXT: v_mov_b32_e32 v23, s16
|
|
; SI-NEXT: v_mov_b32_e32 v24, s17
|
|
; SI-NEXT: v_mov_b32_e32 v21, s18
|
|
; SI-NEXT: v_mov_b32_e32 v22, s19
|
|
; SI-NEXT: v_mov_b32_e32 v19, s20
|
|
; SI-NEXT: v_mov_b32_e32 v20, s21
|
|
; SI-NEXT: v_mov_b32_e32 v17, s22
|
|
; SI-NEXT: v_mov_b32_e32 v18, s23
|
|
; SI-NEXT: v_mov_b32_e32 v15, s24
|
|
; SI-NEXT: v_mov_b32_e32 v16, s25
|
|
; SI-NEXT: v_mov_b32_e32 v13, s26
|
|
; SI-NEXT: v_mov_b32_e32 v14, s27
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: v_mov_b32_e32 v11, s28
|
|
; SI-NEXT: v_mov_b32_e32 v12, s29
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_cbranch_scc0 .LBB49_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_lshr_b64 v[25:26], v[9:10], 16
|
|
; SI-NEXT: v_lshr_b64 v[32:33], v[11:12], 16
|
|
; SI-NEXT: v_lshr_b64 v[26:27], v[7:8], 16
|
|
; SI-NEXT: v_lshr_b64 v[33:34], v[15:16], 16
|
|
; SI-NEXT: v_lshr_b64 v[27:28], v[5:6], 16
|
|
; SI-NEXT: v_lshr_b64 v[34:35], v[17:18], 16
|
|
; SI-NEXT: v_lshr_b64 v[28:29], v[3:4], 16
|
|
; SI-NEXT: v_lshr_b64 v[35:36], v[19:20], 16
|
|
; SI-NEXT: v_lshr_b64 v[29:30], v[1:2], 16
|
|
; SI-NEXT: v_lshr_b64 v[36:37], v[21:22], 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v20
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v22
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v24
|
|
; SI-NEXT: v_lshr_b64 v[30:31], v[13:14], 16
|
|
; SI-NEXT: v_lshr_b64 v[37:38], v[23:24], 16
|
|
; SI-NEXT: s_cbranch_execnz .LBB49_3
|
|
; SI-NEXT: .LBB49_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
|
|
; SI-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
|
|
; SI-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
|
|
; SI-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
|
|
; SI-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
|
|
; SI-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[25:26], v[9:10], 16
|
|
; SI-NEXT: v_lshr_b64 v[32:33], v[11:12], 16
|
|
; SI-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
|
|
; SI-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[26:27], v[7:8], 16
|
|
; SI-NEXT: v_lshr_b64 v[33:34], v[15:16], 16
|
|
; SI-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
|
|
; SI-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[27:28], v[5:6], 16
|
|
; SI-NEXT: v_lshr_b64 v[34:35], v[17:18], 16
|
|
; SI-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
|
|
; SI-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
|
|
; SI-NEXT: v_lshr_b64 v[28:29], v[3:4], 16
|
|
; SI-NEXT: v_lshr_b64 v[35:36], v[19:20], 16
|
|
; SI-NEXT: v_lshr_b64 v[29:30], v[1:2], 16
|
|
; SI-NEXT: v_lshr_b64 v[36:37], v[21:22], 16
|
|
; SI-NEXT: v_lshr_b64 v[30:31], v[13:14], 16
|
|
; SI-NEXT: v_lshr_b64 v[37:38], v[23:24], 16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v18
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v20
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v22
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v24
|
|
; SI-NEXT: .LBB49_3: ; %end
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v37
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v31
|
|
; SI-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v24
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 4, v0
|
|
; SI-NEXT: buffer_store_dword v23, v24, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v23
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 8, v0
|
|
; SI-NEXT: buffer_store_dword v21, v23, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v41
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v22
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 12, v0
|
|
; SI-NEXT: buffer_store_dword v21, v22, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v35
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v21
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 16, v0
|
|
; SI-NEXT: buffer_store_dword v19, v21, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v20
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 20, v0
|
|
; SI-NEXT: buffer_store_dword v19, v20, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v19
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 24, v0
|
|
; SI-NEXT: buffer_store_dword v17, v19, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v18
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 28, v0
|
|
; SI-NEXT: buffer_store_dword v17, v18, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v17
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 32, v0
|
|
; SI-NEXT: buffer_store_dword v15, v17, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v16
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 36, v0
|
|
; SI-NEXT: buffer_store_dword v15, v16, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v30
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v15
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 40, v0
|
|
; SI-NEXT: buffer_store_dword v13, v15, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 44, v0
|
|
; SI-NEXT: buffer_store_dword v13, v14, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v13
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 48, v0
|
|
; SI-NEXT: buffer_store_dword v11, v13, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v52
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v12
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 52, v0
|
|
; SI-NEXT: buffer_store_dword v11, v12, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v29
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v11
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 56, v0
|
|
; SI-NEXT: buffer_store_dword v1, v11, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 60, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v28
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 64, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v50
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v27
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v49
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x4c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x50, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x54, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 0x58, v0
|
|
; SI-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB49_4:
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; 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: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: s_branch .LBB49_2
|
|
;
|
|
; VI-LABEL: bitcast_v12f64_to_v48i16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, s16
|
|
; VI-NEXT: v_mov_b32_e32 v24, s17
|
|
; VI-NEXT: v_mov_b32_e32 v19, s18
|
|
; VI-NEXT: v_mov_b32_e32 v20, s19
|
|
; VI-NEXT: v_mov_b32_e32 v13, s20
|
|
; VI-NEXT: v_mov_b32_e32 v14, s21
|
|
; VI-NEXT: v_mov_b32_e32 v11, s22
|
|
; VI-NEXT: v_mov_b32_e32 v12, s23
|
|
; VI-NEXT: v_mov_b32_e32 v21, s24
|
|
; VI-NEXT: v_mov_b32_e32 v22, s25
|
|
; VI-NEXT: v_mov_b32_e32 v17, s26
|
|
; VI-NEXT: v_mov_b32_e32 v18, s27
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_mov_b32_e32 v15, s28
|
|
; VI-NEXT: v_mov_b32_e32 v16, s29
|
|
; VI-NEXT: s_cbranch_scc0 .LBB49_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; VI-NEXT: s_cbranch_execnz .LBB49_3
|
|
; VI-NEXT: .LBB49_2: ; %cmp.true
|
|
; 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_add_f64 v[15:16], v[15:16], 1.0
|
|
; VI-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
|
|
; VI-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
|
|
; VI-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
|
|
; VI-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
|
|
; VI-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
|
|
; VI-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; VI-NEXT: .LBB49_3: ; %end
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_or_b32_sdwa v30, v23, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v31
|
|
; VI-NEXT: v_or_b32_sdwa v31, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v32
|
|
; VI-NEXT: v_or_b32_sdwa v32, v19, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v33
|
|
; VI-NEXT: v_or_b32_sdwa v33, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v55
|
|
; VI-NEXT: v_or_b32_sdwa v24, v13, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v25
|
|
; VI-NEXT: v_or_b32_sdwa v25, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v26
|
|
; VI-NEXT: v_or_b32_sdwa v26, v11, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v27
|
|
; VI-NEXT: v_lshlrev_b32_e32 v14, 16, v51
|
|
; VI-NEXT: v_or_b32_sdwa v27, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v53
|
|
; VI-NEXT: v_or_b32_sdwa v14, v0, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v12, v15, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v52
|
|
; VI-NEXT: v_or_b32_sdwa v15, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v49
|
|
; VI-NEXT: v_or_b32_sdwa v28, v21, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; VI-NEXT: v_or_b32_sdwa v13, v16, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v48
|
|
; VI-NEXT: v_or_b32_sdwa v29, v22, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v17, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v54
|
|
; VI-NEXT: v_or_b32_sdwa v17, v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v39
|
|
; VI-NEXT: v_or_b32_sdwa v11, v18, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v38
|
|
; VI-NEXT: v_or_b32_sdwa v19, v5, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v37
|
|
; VI-NEXT: v_or_b32_sdwa v20, v6, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v36
|
|
; VI-NEXT: v_or_b32_sdwa v21, v7, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v35
|
|
; VI-NEXT: v_or_b32_sdwa v22, v8, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v34
|
|
; VI-NEXT: v_or_b32_sdwa v23, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, v30
|
|
; VI-NEXT: v_mov_b32_e32 v1, v31
|
|
; VI-NEXT: v_mov_b32_e32 v2, v32
|
|
; VI-NEXT: v_mov_b32_e32 v3, v33
|
|
; VI-NEXT: v_mov_b32_e32 v4, v24
|
|
; VI-NEXT: v_mov_b32_e32 v5, v25
|
|
; VI-NEXT: v_mov_b32_e32 v6, v26
|
|
; VI-NEXT: v_mov_b32_e32 v7, v27
|
|
; VI-NEXT: v_mov_b32_e32 v8, v28
|
|
; VI-NEXT: v_mov_b32_e32 v9, v29
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB49_4:
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; VI-NEXT: ; implicit-def: $vgpr10
|
|
; 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: s_branch .LBB49_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v12f64_to_v48i16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s27
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s29
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB49_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB49_3
|
|
; GFX9-NEXT: .LBB49_2: ; %cmp.true
|
|
; 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_add_f64 v[15:16], v[15:16], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; GFX9-NEXT: .LBB49_3: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v30, v30, 16, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v24
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v55, 16, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v26, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v27, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v21
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v50, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v25, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v28, v28, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v22
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v49, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v3
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v29, v29, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v48, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v32, v32, 16, v19
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v10, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v39, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v33, v33, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v38, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v37, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v7
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v36, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v31, v31, 16, v23
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v35, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v54, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v53, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v52, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v34, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, v30
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, v31
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, v32
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, v25
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, v26
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, v27
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, v28
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, v29
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB49_4:
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; GFX9-NEXT: ; implicit-def: $vgpr10
|
|
; 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: s_branch .LBB49_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v12f64_to_v48i16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB49_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB49_3
|
|
; GFX11-TRUE16-NEXT: .LBB49_2: ; %cmp.true
|
|
; 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: v_lshrrev_b32_e32 v24, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; GFX11-TRUE16-NEXT: .LBB49_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v55, v55 :: v_dual_mov_b32 v54, v54
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v53 :: v_dual_mov_b32 v52, v52
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v51, v51 :: v_dual_mov_b32 v50, v50
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v49, v49 :: v_dual_mov_b32 v48, v48
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, v39 :: v_dual_mov_b32 v38, v38
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, v37 :: v_dual_mov_b32 v36, v36
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, v35 :: v_dual_mov_b32 v34, v34
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, v33 :: v_dual_mov_b32 v32, v32
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, v31 :: v_dual_mov_b32 v30, v30
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v29 :: v_dual_mov_b32 v28, v28
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v27 :: v_dual_mov_b32 v26, v26
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v25 :: v_dual_mov_b32 v24, v24
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v31.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v28.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v27.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v26.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v25.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v24.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB49_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB49_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v12f64_to_v48i16_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v25, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v23, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v21, s17
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v19, s21
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s24 :: v_dual_mov_b32 v10, s25
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB49_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v13, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v6, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB49_3
|
|
; GFX11-FAKE16-NEXT: .LBB49_2: ; %cmp.true
|
|
; 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_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[9:10], v[9:10], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v13, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v6, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: .LBB49_3: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v29, v29, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v6, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v53, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v52, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v54, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v27, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v51, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v28, v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v13, 16, v20
|
|
; 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 v17, 0xffff, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v33, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v49, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v48, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v39, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v38, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v37, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v36, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v34, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v32, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v31, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v30, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v0, v24
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v28
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB49_4:
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr6
|
|
; 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: $vgpr13
|
|
; 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: s_branch .LBB49_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <12 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <12 x double> %a1 to <48 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x double> %a to <48 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x i16> %phi
|
|
}
|
|
|
|
define <12 x double> @bitcast_v48i16_to_v12f64(<48 x i16> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v48i16_to_v12f64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v48, v14
|
|
; SI-NEXT: v_mov_b32_e32 v49, v12
|
|
; SI-NEXT: v_mov_b32_e32 v50, v10
|
|
; SI-NEXT: v_mov_b32_e32 v51, v8
|
|
; SI-NEXT: v_mov_b32_e32 v52, v6
|
|
; SI-NEXT: v_mov_b32_e32 v53, v4
|
|
; SI-NEXT: v_mov_b32_e32 v54, v2
|
|
; SI-NEXT: v_mov_b32_e32 v55, v0
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:68
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:60
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_lshlrev_b32_e32 v47, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v39, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v37, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v36, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v43, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v35, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v42, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v41, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v33, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v29
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v61, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v8
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v12
|
|
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:56
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:52
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:44
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:40
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:36
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:28
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:20
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:196 ; 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 v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v55
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v54
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v53
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v52
|
|
; SI-NEXT: v_and_b32_e32 v4, 0xffff, v51
|
|
; SI-NEXT: v_and_b32_e32 v5, 0xffff, v50
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v49
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v48
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_or_b32_e32 v0, v0, v47
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v39
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v46
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v38
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v45
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v37
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v44
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v36
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; 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: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_and_b32_e32 v10, 0xffff, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v43
|
|
; SI-NEXT: v_or_b32_e32 v9, v9, v35
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v42
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v34
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v41
|
|
; SI-NEXT: v_or_b32_e32 v13, v13, v33
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v40
|
|
; SI-NEXT: v_or_b32_e32 v15, v15, v32
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v63
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v62
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v61
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v60
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v59
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v58
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v57
|
|
; SI-NEXT: v_or_b32_e32 v23, v23, v56
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; 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 v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v55
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v54
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v53
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v52
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v51
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v50
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v49
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v48
|
|
; 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_or_b32_e32 v0, v47, v0
|
|
; SI-NEXT: s_mov_b32 s6, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v1, v39, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v46, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v38, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v45, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v37, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v44, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v36, v7
|
|
; 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: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 3, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; 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_and_b32_e32 v12, 0xffff, v12
|
|
; SI-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; SI-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; SI-NEXT: v_and_b32_e32 v15, 0xffff, v15
|
|
; SI-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; SI-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; SI-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; SI-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; SI-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; SI-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; SI-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; SI-NEXT: v_or_b32_e32 v8, v43, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v35, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v42, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v34, v11
|
|
; SI-NEXT: v_or_b32_e32 v12, v41, v12
|
|
; SI-NEXT: v_or_b32_e32 v13, v33, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v40, v14
|
|
; SI-NEXT: v_or_b32_e32 v15, v32, v15
|
|
; SI-NEXT: v_or_b32_e32 v16, v63, v16
|
|
; SI-NEXT: v_or_b32_e32 v17, v62, v17
|
|
; SI-NEXT: v_or_b32_e32 v18, v61, v18
|
|
; SI-NEXT: v_or_b32_e32 v19, v60, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v59, v20
|
|
; SI-NEXT: v_or_b32_e32 v21, v58, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v57, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v56, v23
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, s6, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, s6, v9
|
|
; 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: v_add_i32_e32 v20, vcc, s6, v20
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, s6, v21
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, s6, v22
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 0x30000, v23
|
|
; 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 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v48i16_to_v12f64:
|
|
; 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: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_mov_b32_e32 v32, v23
|
|
; VI-NEXT: v_mov_b32_e32 v33, v22
|
|
; VI-NEXT: v_mov_b32_e32 v34, v21
|
|
; VI-NEXT: v_mov_b32_e32 v35, v20
|
|
; VI-NEXT: v_mov_b32_e32 v36, v19
|
|
; VI-NEXT: v_mov_b32_e32 v37, v18
|
|
; VI-NEXT: v_mov_b32_e32 v38, v17
|
|
; VI-NEXT: v_mov_b32_e32 v39, v16
|
|
; VI-NEXT: v_mov_b32_e32 v48, v15
|
|
; VI-NEXT: v_mov_b32_e32 v49, v14
|
|
; VI-NEXT: v_mov_b32_e32 v50, v13
|
|
; VI-NEXT: v_mov_b32_e32 v51, v12
|
|
; VI-NEXT: v_mov_b32_e32 v52, v11
|
|
; VI-NEXT: v_mov_b32_e32 v53, v10
|
|
; VI-NEXT: v_mov_b32_e32 v54, v9
|
|
; VI-NEXT: v_mov_b32_e32 v55, v8
|
|
; VI-NEXT: v_mov_b32_e32 v40, v7
|
|
; VI-NEXT: v_mov_b32_e32 v41, v6
|
|
; VI-NEXT: v_mov_b32_e32 v42, v5
|
|
; VI-NEXT: v_mov_b32_e32 v43, v4
|
|
; VI-NEXT: v_mov_b32_e32 v44, v3
|
|
; VI-NEXT: v_mov_b32_e32 v45, v2
|
|
; VI-NEXT: v_mov_b32_e32 v46, v1
|
|
; VI-NEXT: v_mov_b32_e32 v47, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 v23, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v23, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v23, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v23, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v23, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v23, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v23, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v23, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v23, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v23, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v23, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v23, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v23, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v23, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v23, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v23, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v23, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v23, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v23, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v23, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v23, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v23, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v23, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v23, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v44, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v43, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v42, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v41, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v40, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v55, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v53, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v52, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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 v23, 3
|
|
; VI-NEXT: v_add_u16_e32 v0, 3, v47
|
|
; VI-NEXT: v_add_u16_sdwa v1, v47, v23 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_add_u16_e32 v1, 3, v46
|
|
; VI-NEXT: v_add_u16_sdwa v2, v46, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v45
|
|
; VI-NEXT: v_add_u16_sdwa v3, v45, v23 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, v44
|
|
; VI-NEXT: v_add_u16_sdwa v4, v44, v23 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, v43
|
|
; VI-NEXT: v_add_u16_sdwa v5, v43, v23 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, v42
|
|
; VI-NEXT: v_add_u16_sdwa v6, v42, v23 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, v41
|
|
; VI-NEXT: v_add_u16_sdwa v7, v41, v23 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, v40
|
|
; VI-NEXT: v_add_u16_sdwa v8, v40, v23 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, v55
|
|
; VI-NEXT: v_add_u16_sdwa v9, v55, v23 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, v54
|
|
; VI-NEXT: v_add_u16_sdwa v10, v54, v23 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, v53
|
|
; VI-NEXT: v_add_u16_sdwa v11, v53, v23 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, v52
|
|
; VI-NEXT: v_add_u16_sdwa v12, v52, v23 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, v51
|
|
; VI-NEXT: v_add_u16_sdwa v13, v51, v23 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, v50
|
|
; VI-NEXT: v_add_u16_sdwa v14, v50, v23 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, v49
|
|
; VI-NEXT: v_add_u16_sdwa v15, v49, v23 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, v48
|
|
; VI-NEXT: v_add_u16_sdwa v16, v48, v23 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, v39
|
|
; VI-NEXT: v_add_u16_sdwa v17, v39, v23 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, v38
|
|
; VI-NEXT: v_add_u16_sdwa v18, v38, v23 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, v37
|
|
; VI-NEXT: v_add_u16_sdwa v19, v37, v23 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, v36
|
|
; VI-NEXT: v_add_u16_sdwa v20, v36, v23 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, v35
|
|
; VI-NEXT: v_add_u16_sdwa v21, v35, v23 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, v34
|
|
; VI-NEXT: v_add_u16_sdwa v22, v34, v23 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, v33
|
|
; VI-NEXT: v_add_u16_sdwa v24, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v22, v22, v24
|
|
; VI-NEXT: v_add_u16_e32 v24, 3, v32
|
|
; VI-NEXT: v_add_u16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB50_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48i16_to_v12f64:
|
|
; 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 v33, v22
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v13
|
|
; 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, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v12
|
|
; 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, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v11
|
|
; 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, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v10
|
|
; 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, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v1
|
|
; 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, v53
|
|
; 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, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v47
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 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:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; 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: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; 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 v9, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, 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(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, 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: .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_v48i16_to_v12f64:
|
|
; 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, v24
|
|
; 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: .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_v48i16_to_v12f64:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v24
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v53, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v55, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v64, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v52, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v51, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v50, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v49, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v48, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v39, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v38, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v37, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v36, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v35, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v34, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v33, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v32, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v31, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v30, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v29, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v28, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v27, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v26, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v25, v23, 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: .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 <48 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <48 x i16> %a1 to <12 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x i16> %a to <12 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x double> %phi
|
|
}
|
|
|
|
define inreg <12 x double> @bitcast_v48i16_to_v12f64_scalar(<48 x i16> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v48i16_to_v12f64_scalar:
|
|
; 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:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v61, v4
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v63, v2
|
|
; SI-NEXT: v_mov_b32_e32 v56, v0
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: v_mov_b32_e32 v31, v22
|
|
; SI-NEXT: v_mov_b32_e32 v34, v20
|
|
; SI-NEXT: v_mov_b32_e32 v35, v18
|
|
; SI-NEXT: v_mov_b32_e32 v36, v16
|
|
; SI-NEXT: v_mov_b32_e32 v37, v14
|
|
; SI-NEXT: v_mov_b32_e32 v38, v12
|
|
; SI-NEXT: v_mov_b32_e32 v39, v10
|
|
; SI-NEXT: v_mov_b32_e32 v48, v8
|
|
; SI-NEXT: v_mov_b32_e32 v50, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v54, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v57, 16, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v55, 16, v7
|
|
; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v60, 16, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v59, 16, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v58, 16, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v41, 16, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v23
|
|
; SI-NEXT: v_lshlrev_b32_e32 v52, 16, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v51, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v2
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v4
|
|
; SI-NEXT: s_cbranch_scc0 .LBB51_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v56
|
|
; SI-NEXT: v_or_b32_e32 v7, v0, v54
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v61
|
|
; SI-NEXT: v_or_b32_e32 v9, v0, v32
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v50
|
|
; SI-NEXT: v_or_b32_e32 v10, v0, v55
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; SI-NEXT: v_or_b32_e32 v11, v0, v62
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; SI-NEXT: v_or_b32_e32 v12, v0, v46
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; SI-NEXT: v_or_b32_e32 v13, v0, v45
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; SI-NEXT: v_or_b32_e32 v14, v0, v60
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; SI-NEXT: v_or_b32_e32 v15, v0, v59
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s17, 16
|
|
; SI-NEXT: v_or_b32_e32 v16, v0, v58
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; SI-NEXT: s_or_b32 s4, s4, s5
|
|
; SI-NEXT: s_and_b32 s5, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s19, 16
|
|
; SI-NEXT: v_or_b32_e32 v17, v0, v41
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v31
|
|
; SI-NEXT: s_or_b32 s5, s5, s6
|
|
; SI-NEXT: s_and_b32 s6, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s21, 16
|
|
; SI-NEXT: v_or_b32_e32 v18, v0, v40
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v24
|
|
; SI-NEXT: s_or_b32 s6, s6, s7
|
|
; SI-NEXT: s_and_b32 s7, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s23, 16
|
|
; SI-NEXT: v_or_b32_e32 v19, v0, v52
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v26
|
|
; SI-NEXT: s_or_b32 s7, s7, s8
|
|
; SI-NEXT: s_and_b32 s8, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s25, 16
|
|
; SI-NEXT: v_or_b32_e32 v20, v0, v51
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v28
|
|
; SI-NEXT: s_or_b32 s8, s8, s9
|
|
; SI-NEXT: s_and_b32 s9, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s10, s27, 16
|
|
; SI-NEXT: v_or_b32_e32 v21, v0, v29
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v30
|
|
; SI-NEXT: s_or_b32 s9, s9, s10
|
|
; SI-NEXT: s_and_b32 s10, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s11, s29, 16
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v63
|
|
; SI-NEXT: v_or_b32_e32 v22, v0, v27
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; SI-NEXT: s_or_b32 s10, s10, s11
|
|
; SI-NEXT: v_or_b32_e32 v8, v1, v57
|
|
; SI-NEXT: v_or_b32_e32 v23, v0, v25
|
|
; SI-NEXT: v_mov_b32_e32 v0, s4
|
|
; SI-NEXT: v_mov_b32_e32 v1, s5
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: v_mov_b32_e32 v3, s7
|
|
; SI-NEXT: v_mov_b32_e32 v4, s8
|
|
; SI-NEXT: v_mov_b32_e32 v5, s9
|
|
; SI-NEXT: v_mov_b32_e32 v6, s10
|
|
; SI-NEXT: s_cbranch_execnz .LBB51_3
|
|
; SI-NEXT: .LBB51_2: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v56
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v54, v0
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v61
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v32, v0
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v50
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_or_b32_e32 v0, v55, v0
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v48
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v62, v0
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v39
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_or_b32_e32 v0, v46, v0
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v38
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v45, v0
|
|
; SI-NEXT: v_add_i32_e32 v13, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v37
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v60, v0
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v36
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v59, v0
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v35
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v58, v0
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v34
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v41, v0
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v31
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v40, v0
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v24
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v0, v52, v0
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v26
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_add_i32 s16, s16, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v51, v0
|
|
; SI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s5, s17, 16
|
|
; SI-NEXT: s_add_i32 s18, s18, 3
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v28
|
|
; SI-NEXT: s_or_b32 s4, s5, s4
|
|
; SI-NEXT: s_and_b32 s5, s18, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s6, s19, 16
|
|
; SI-NEXT: s_add_i32 s20, s20, 3
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_or_b32 s5, s6, s5
|
|
; SI-NEXT: s_and_b32 s6, s20, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s7, s21, 16
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v29, v0
|
|
; SI-NEXT: s_or_b32 s6, s7, s6
|
|
; SI-NEXT: s_and_b32 s7, s22, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s8, s23, 16
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v30
|
|
; SI-NEXT: s_or_b32 s7, s8, s7
|
|
; SI-NEXT: s_and_b32 s8, s24, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s9, s25, 16
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: s_or_b32 s8, s9, s8
|
|
; SI-NEXT: s_and_b32 s9, s26, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s10, s27, 16
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: v_or_b32_e32 v0, v27, v0
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v63
|
|
; SI-NEXT: s_or_b32 s9, s10, s9
|
|
; SI-NEXT: s_and_b32 s10, s28, 0xffff
|
|
; SI-NEXT: s_lshl_b32 s11, s29, 16
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 0x30000, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v49
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; SI-NEXT: s_or_b32 s10, s11, s10
|
|
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v57, v1
|
|
; SI-NEXT: s_add_i32 s4, s4, 0x30000
|
|
; SI-NEXT: s_add_i32 s5, s5, 0x30000
|
|
; SI-NEXT: s_add_i32 s6, s6, 0x30000
|
|
; SI-NEXT: s_add_i32 s7, s7, 0x30000
|
|
; SI-NEXT: s_add_i32 s8, s8, 0x30000
|
|
; SI-NEXT: s_add_i32 s9, s9, 0x30000
|
|
; SI-NEXT: s_add_i32 s10, s10, 0x30000
|
|
; SI-NEXT: v_or_b32_e32 v0, v25, v0
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 0x30000, v1
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 0x30000, v0
|
|
; SI-NEXT: v_mov_b32_e32 v0, s4
|
|
; SI-NEXT: v_mov_b32_e32 v1, s5
|
|
; SI-NEXT: v_mov_b32_e32 v2, s6
|
|
; SI-NEXT: v_mov_b32_e32 v3, s7
|
|
; SI-NEXT: v_mov_b32_e32 v4, s8
|
|
; SI-NEXT: v_mov_b32_e32 v5, s9
|
|
; SI-NEXT: v_mov_b32_e32 v6, s10
|
|
; SI-NEXT: .LBB51_3: ; %end
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB51_4:
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v46, v51
|
|
; SI-NEXT: v_mov_b32_e32 v51, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v34
|
|
; SI-NEXT: v_mov_b32_e32 v34, v30
|
|
; SI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_mov_b32_e32 v47, v52
|
|
; SI-NEXT: v_mov_b32_e32 v52, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v35
|
|
; SI-NEXT: v_mov_b32_e32 v35, v28
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v55, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v41
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v41, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v26
|
|
; SI-NEXT: v_mov_b32_e32 v42, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v24
|
|
; SI-NEXT: v_mov_b32_e32 v33, v32
|
|
; SI-NEXT: v_mov_b32_e32 v32, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v40
|
|
; SI-NEXT: v_mov_b32_e32 v40, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v31
|
|
; SI-NEXT: v_mov_b32_e32 v43, v25
|
|
; SI-NEXT: v_mov_b32_e32 v44, v27
|
|
; SI-NEXT: v_mov_b32_e32 v53, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v29
|
|
; 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: v_mov_b32_e32 v26, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v55
|
|
; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v30, v34
|
|
; SI-NEXT: v_mov_b32_e32 v34, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v51
|
|
; SI-NEXT: v_mov_b32_e32 v51, v46
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v29, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v53
|
|
; SI-NEXT: v_mov_b32_e32 v27, v44
|
|
; SI-NEXT: v_mov_b32_e32 v25, v43
|
|
; SI-NEXT: v_mov_b32_e32 v31, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v40
|
|
; SI-NEXT: v_mov_b32_e32 v40, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v32
|
|
; SI-NEXT: v_mov_b32_e32 v32, v33
|
|
; SI-NEXT: v_mov_b32_e32 v24, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v42
|
|
; SI-NEXT: v_mov_b32_e32 v28, v35
|
|
; SI-NEXT: v_mov_b32_e32 v35, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v52
|
|
; SI-NEXT: v_mov_b32_e32 v52, v47
|
|
; SI-NEXT: s_branch .LBB51_2
|
|
;
|
|
; VI-LABEL: bitcast_v48i16_to_v12f64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v32, v9
|
|
; VI-NEXT: v_mov_b32_e32 v33, v8
|
|
; VI-NEXT: v_mov_b32_e32 v34, v7
|
|
; VI-NEXT: v_mov_b32_e32 v35, v6
|
|
; VI-NEXT: v_mov_b32_e32 v36, v5
|
|
; VI-NEXT: v_mov_b32_e32 v37, v4
|
|
; VI-NEXT: v_mov_b32_e32 v38, v3
|
|
; VI-NEXT: v_mov_b32_e32 v39, v2
|
|
; VI-NEXT: v_mov_b32_e32 v48, v1
|
|
; VI-NEXT: v_mov_b32_e32 v49, v0
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; 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, s43, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s44, s42, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s44
|
|
; VI-NEXT: s_and_b32 s44, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s45, s41, 16
|
|
; VI-NEXT: s_or_b32 s44, s44, s45
|
|
; VI-NEXT: s_and_b32 s45, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s46, s40, 16
|
|
; VI-NEXT: s_or_b32 s45, s45, s46
|
|
; VI-NEXT: s_and_b32 s46, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s47, s15, 16
|
|
; VI-NEXT: s_or_b32 s46, s46, s47
|
|
; VI-NEXT: s_and_b32 s47, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s56, s14, 16
|
|
; VI-NEXT: v_mov_b32_e32 v0, 16
|
|
; VI-NEXT: s_or_b32 s47, s47, s56
|
|
; VI-NEXT: s_and_b32 s56, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s57, s13, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s56, s56, s57
|
|
; VI-NEXT: s_and_b32 s57, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s58, s12, 16
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s57, s57, s58
|
|
; VI-NEXT: s_and_b32 s58, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s59, s11, 16
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s58, s58, s59
|
|
; VI-NEXT: s_and_b32 s59, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s60, s10, 16
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s59, s59, s60
|
|
; VI-NEXT: s_and_b32 s60, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s61, s9, 16
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s60, s60, s61
|
|
; VI-NEXT: s_and_b32 s61, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s62, s8, 16
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s61, s61, s62
|
|
; VI-NEXT: s_and_b32 s62, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s63, s7, 16
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s62, s62, s63
|
|
; VI-NEXT: s_and_b32 s63, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s72, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s63, s63, s72
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v0, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s44
|
|
; VI-NEXT: v_mov_b32_e32 v3, s45
|
|
; VI-NEXT: v_mov_b32_e32 v4, s46
|
|
; VI-NEXT: v_mov_b32_e32 v5, s47
|
|
; VI-NEXT: v_mov_b32_e32 v6, s56
|
|
; VI-NEXT: v_mov_b32_e32 v7, s57
|
|
; VI-NEXT: v_mov_b32_e32 v8, s58
|
|
; VI-NEXT: v_mov_b32_e32 v9, s59
|
|
; VI-NEXT: v_mov_b32_e32 v10, s60
|
|
; VI-NEXT: v_mov_b32_e32 v11, s61
|
|
; VI-NEXT: v_mov_b32_e32 v12, s62
|
|
; VI-NEXT: v_mov_b32_e32 v13, s63
|
|
; VI-NEXT: s_cbranch_execnz .LBB51_3
|
|
; VI-NEXT: .LBB51_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v1, 16
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v49
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: v_add_u32_e32 v14, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v48
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_add_i32 s16, s16, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_and_b32 s4, s16, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s5, s43, 16
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: v_add_u32_e32 v15, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v39
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s4, s5, s4
|
|
; VI-NEXT: s_and_b32 s5, s17, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s16, s42, 16
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s5, s16, s5
|
|
; VI-NEXT: s_and_b32 s16, s18, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s17, s41, 16
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v38
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s16, s17, s16
|
|
; VI-NEXT: s_and_b32 s17, s19, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s18, s40, 16
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s17, s18, s17
|
|
; VI-NEXT: s_and_b32 s18, s20, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s15, s15, 16
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: v_add_u32_e32 v17, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v37
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s15, s15, s18
|
|
; VI-NEXT: s_and_b32 s18, s21, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s14, s14, 16
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s14, s14, s18
|
|
; VI-NEXT: s_and_b32 s18, s22, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s13, s13, 16
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v36
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s13, s13, s18
|
|
; VI-NEXT: s_and_b32 s18, s23, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s12, s12, 16
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s12, s12, s18
|
|
; VI-NEXT: s_and_b32 s18, s24, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s11, s11, 16
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: v_add_u32_e32 v19, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v35
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s11, s11, s18
|
|
; VI-NEXT: s_and_b32 s18, s25, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s10, s10, 16
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s10, s10, s18
|
|
; VI-NEXT: s_and_b32 s18, s26, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s9, s9, 16
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v34
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s9, s9, s18
|
|
; VI-NEXT: s_and_b32 s18, s27, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s8, s8, 16
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s8, s8, s18
|
|
; VI-NEXT: s_and_b32 s18, s28, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s7, s7, 16
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: v_add_u32_e32 v21, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v33
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v1, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s7, s7, s18
|
|
; VI-NEXT: s_and_b32 s18, s29, 0xffff
|
|
; VI-NEXT: s_lshl_b32 s6, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: s_or_b32 s6, s6, s18
|
|
; VI-NEXT: v_add_u32_e32 v22, vcc, 0x30000, v0
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v32
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v1, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_add_i32 s4, s4, 0x30000
|
|
; VI-NEXT: s_add_i32 s5, s5, 0x30000
|
|
; VI-NEXT: s_add_i32 s16, s16, 0x30000
|
|
; VI-NEXT: s_add_i32 s17, s17, 0x30000
|
|
; VI-NEXT: s_add_i32 s15, s15, 0x30000
|
|
; VI-NEXT: s_add_i32 s14, s14, 0x30000
|
|
; VI-NEXT: s_add_i32 s13, s13, 0x30000
|
|
; VI-NEXT: s_add_i32 s12, s12, 0x30000
|
|
; VI-NEXT: s_add_i32 s11, s11, 0x30000
|
|
; VI-NEXT: s_add_i32 s10, s10, 0x30000
|
|
; VI-NEXT: s_add_i32 s9, s9, 0x30000
|
|
; VI-NEXT: s_add_i32 s8, s8, 0x30000
|
|
; VI-NEXT: s_add_i32 s7, s7, 0x30000
|
|
; VI-NEXT: s_add_i32 s6, s6, 0x30000
|
|
; VI-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
|
|
; VI-NEXT: v_add_u32_e32 v23, vcc, 0x30000, v0
|
|
; 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, s15
|
|
; VI-NEXT: v_mov_b32_e32 v5, s14
|
|
; VI-NEXT: v_mov_b32_e32 v6, s13
|
|
; VI-NEXT: v_mov_b32_e32 v7, s12
|
|
; VI-NEXT: v_mov_b32_e32 v8, s11
|
|
; VI-NEXT: v_mov_b32_e32 v9, s10
|
|
; VI-NEXT: v_mov_b32_e32 v10, s9
|
|
; VI-NEXT: v_mov_b32_e32 v11, s8
|
|
; VI-NEXT: v_mov_b32_e32 v12, s7
|
|
; VI-NEXT: v_mov_b32_e32 v13, s6
|
|
; VI-NEXT: .LBB51_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB51_4:
|
|
; 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_branch .LBB51_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v48i16_to_v12f64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v0
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, 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: buffer_store_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v32
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v33
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v49
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s17, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s18, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s19, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s20, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s21, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s22, s12
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s23, s13
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s24, s14
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s25, s15
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s29, s40
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB51_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v33
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s19
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB51_3
|
|
; GFX9-NEXT: .LBB51_2: ; %cmp.true
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v49
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v48
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v39
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v38
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v37
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v36
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v35
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v34
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v33
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v23
|
|
; GFX9-NEXT: v_pk_add_u16 v0, s6, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v1, s7, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v2, s8, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v3, s9, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v4, s10, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v5, s11, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v6, s12, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v7, s13, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v8, s14, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v9, s15, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v10, s16, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v11, s17, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v12, s18, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_u16 v13, s19, 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: .LBB51_3: ; %end
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB51_4:
|
|
; 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_branch .LBB51_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v12f64_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB51_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; 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, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB51_3
|
|
; GFX11-TRUE16-NEXT: .LBB51_2: ; %cmp.true
|
|
; 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, s4, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s17, 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: .LBB51_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB51_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB51_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v48i16_to_v12f64_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s4, s0, s4
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s44
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s16, s5
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s17, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s18, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s19, s8
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s20, s9
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s21, s10
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s22, s11
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s23, s12
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s24, s13
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s25, s14
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s26, s43
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB51_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB51_3
|
|
; GFX11-FAKE16-NEXT: .LBB51_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v0, s4, 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 v2, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, s17, 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: .LBB51_3: ; %end
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB51_4:
|
|
; GFX11-FAKE16-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
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB51_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = add <48 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <48 x i16> %a1 to <12 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x i16> %a to <12 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x double> %phi
|
|
}
|
|
|
|
define <48 x half> @bitcast_v12f64_to_v48f16(<12 x double> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v12f64_to_v48f16:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v25
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; 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: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; 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: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; 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_lshrrev_b32_e32 v25, 16, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v1
|
|
; SI-NEXT: ; implicit-def: $vgpr1
|
|
; SI-NEXT: ; implicit-def: $vgpr3
|
|
; SI-NEXT: ; implicit-def: $vgpr5
|
|
; SI-NEXT: ; implicit-def: $vgpr7
|
|
; SI-NEXT: ; implicit-def: $vgpr9
|
|
; SI-NEXT: ; implicit-def: $vgpr11
|
|
; SI-NEXT: ; implicit-def: $vgpr13
|
|
; 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: .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[1:2], v[1:2], 1.0
|
|
; SI-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
|
|
; SI-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
|
|
; SI-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
|
|
; SI-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
|
|
; SI-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
|
|
; SI-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
|
|
; SI-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
|
|
; SI-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
|
|
; SI-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
|
|
; SI-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
|
|
; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v1
|
|
; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v3
|
|
; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v5
|
|
; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v7
|
|
; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v9
|
|
; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v10
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v11
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v13
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v15
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v21
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v51
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v49
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v36
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v48
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v52
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v40
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v45
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v47
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v59
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v61
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v33
|
|
; SI-NEXT: v_mov_b32_e32 v35, v24
|
|
; SI-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: .LBB52_4: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v30
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v3
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
|
|
; SI-NEXT: v_or_b32_e32 v2, v4, v2
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v29
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v28
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v26
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v63
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v25
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v61
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v62
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v59
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v60
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v57
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v58
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v56
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v46
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v43
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v44
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v42
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v41
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v55
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v50
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v53
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v48
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v38
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v36
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v34
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v37
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v39
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v49
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v51
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; 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_v12f64_to_v48f16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 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[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 v24, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; VI-NEXT: .LBB52_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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_lshlrev_b32_e32 v27, 16, v27
|
|
; VI-NEXT: v_lshlrev_b32_e32 v26, 16, v26
|
|
; VI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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_or_b32_sdwa v20, v20, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v22, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v12f64_to_v48f16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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[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 v24, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v54, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v53, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v52, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v51, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v50, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v49, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v48, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v39, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v38, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v37, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v36, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v35, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v34, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v33, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v32, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v31, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v30, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v29, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v28, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v27, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v26, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v25, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s4
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v12f64_to_v48f16:
|
|
; 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, v24
|
|
; 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[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_v12f64_to_v48f16:
|
|
; 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, v24
|
|
; 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: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr24
|
|
; 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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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[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 v24, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 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, v55, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v53, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v52, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v51, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v50, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v49, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v48, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v39, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v38, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v37, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v35, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v34, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v33, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v32, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v31, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v30, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v29, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v28, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v27, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v26, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v25, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v24, v23, 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 <12 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <12 x double> %a1 to <48 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x double> %a to <48 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x half> %phi
|
|
}
|
|
|
|
define inreg <48 x half> @bitcast_v12f64_to_v48f16_scalar(<12 x double> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v12f64_to_v48f16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v11
|
|
; SI-NEXT: v_mov_b32_e32 v23, s16
|
|
; SI-NEXT: v_mov_b32_e32 v24, s17
|
|
; SI-NEXT: v_mov_b32_e32 v21, s18
|
|
; SI-NEXT: v_mov_b32_e32 v22, s19
|
|
; SI-NEXT: v_mov_b32_e32 v19, s20
|
|
; SI-NEXT: v_mov_b32_e32 v20, s21
|
|
; SI-NEXT: v_mov_b32_e32 v15, s22
|
|
; SI-NEXT: v_mov_b32_e32 v16, s23
|
|
; SI-NEXT: v_mov_b32_e32 v17, s24
|
|
; SI-NEXT: v_mov_b32_e32 v18, s25
|
|
; SI-NEXT: v_mov_b32_e32 v13, s26
|
|
; SI-NEXT: v_mov_b32_e32 v14, s27
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: v_mov_b32_e32 v11, s28
|
|
; SI-NEXT: v_mov_b32_e32 v12, s29
|
|
; 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 .LBB53_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v26
|
|
; SI-NEXT: v_lshrrev_b32_e32 v26, 16, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v9
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v9
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v25
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v7
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, v10
|
|
; SI-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v4
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v12
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v18
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v20
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, v25
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v15
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v23
|
|
; SI-NEXT: s_cbranch_execnz .LBB53_3
|
|
; SI-NEXT: .LBB53_2: ; %cmp.true
|
|
; SI-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
|
|
; SI-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
|
|
; SI-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
|
|
; SI-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
|
|
; SI-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
|
|
; SI-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
|
|
; SI-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
|
|
; SI-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
|
|
; SI-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
|
|
; SI-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
|
|
; SI-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
|
|
; SI-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
|
|
; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v24
|
|
; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v21
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v22
|
|
; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v19
|
|
; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v20
|
|
; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v15
|
|
; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v16
|
|
; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v17
|
|
; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v18
|
|
; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v13
|
|
; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v14
|
|
; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v11
|
|
; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v12
|
|
; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v2
|
|
; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v3
|
|
; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v4
|
|
; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v5
|
|
; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v6
|
|
; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v7
|
|
; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v9
|
|
; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v55, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v41, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, v16
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v51
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v49
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v36
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v48
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v52
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v54, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v40, v40
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, v42
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v44
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, v46
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v56
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v58
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, v60
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, v62
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v28
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v30
|
|
; SI-NEXT: v_mov_b32_e32 v35, v10
|
|
; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: .LBB53_3: ; %end
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v3
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
|
|
; SI-NEXT: v_or_b32_e32 v2, v4, v2
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v25
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v31
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v62
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v29
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v60
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v27
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v58
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v26
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v56
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v63
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v46
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v61
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v44
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v59
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v43
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v57
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v42
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v47
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v45
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v41
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v55
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v50
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v53
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v48
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v38
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v36
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v34
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v37
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v39
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v49
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v51
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; 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]
|
|
; SI-NEXT: .LBB53_4:
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; kill: killed $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; kill: killed $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: s_branch .LBB53_2
|
|
;
|
|
; VI-LABEL: bitcast_v12f64_to_v48f16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v23, s16
|
|
; VI-NEXT: v_mov_b32_e32 v24, s17
|
|
; VI-NEXT: v_mov_b32_e32 v19, s18
|
|
; VI-NEXT: v_mov_b32_e32 v20, s19
|
|
; VI-NEXT: v_mov_b32_e32 v13, s20
|
|
; VI-NEXT: v_mov_b32_e32 v14, s21
|
|
; VI-NEXT: v_mov_b32_e32 v11, s22
|
|
; VI-NEXT: v_mov_b32_e32 v12, s23
|
|
; VI-NEXT: v_mov_b32_e32 v21, s24
|
|
; VI-NEXT: v_mov_b32_e32 v22, s25
|
|
; VI-NEXT: v_mov_b32_e32 v17, s26
|
|
; VI-NEXT: v_mov_b32_e32 v18, s27
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_mov_b32_e32 v15, s28
|
|
; VI-NEXT: v_mov_b32_e32 v16, s29
|
|
; VI-NEXT: s_cbranch_scc0 .LBB53_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; VI-NEXT: s_cbranch_execnz .LBB53_3
|
|
; VI-NEXT: .LBB53_2: ; %cmp.true
|
|
; 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_add_f64 v[15:16], v[15:16], 1.0
|
|
; VI-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
|
|
; VI-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
|
|
; VI-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
|
|
; VI-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
|
|
; VI-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
|
|
; VI-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; VI-NEXT: .LBB53_3: ; %end
|
|
; VI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; VI-NEXT: v_or_b32_sdwa v30, v23, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v31
|
|
; VI-NEXT: v_or_b32_sdwa v31, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v23, 16, v32
|
|
; VI-NEXT: v_or_b32_sdwa v32, v19, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v33
|
|
; VI-NEXT: v_or_b32_sdwa v33, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v19, 16, v55
|
|
; VI-NEXT: v_or_b32_sdwa v24, v13, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v25
|
|
; VI-NEXT: v_or_b32_sdwa v25, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v26
|
|
; VI-NEXT: v_or_b32_sdwa v26, v11, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v27
|
|
; VI-NEXT: v_lshlrev_b32_e32 v14, 16, v51
|
|
; VI-NEXT: v_or_b32_sdwa v27, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v12, 16, v53
|
|
; VI-NEXT: v_or_b32_sdwa v14, v0, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v50
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v12, v15, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v13, 16, v52
|
|
; VI-NEXT: v_or_b32_sdwa v15, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v49
|
|
; VI-NEXT: v_or_b32_sdwa v28, v21, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v29
|
|
; VI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; VI-NEXT: v_or_b32_sdwa v13, v16, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v48
|
|
; VI-NEXT: v_or_b32_sdwa v29, v22, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v17, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v11, 16, v54
|
|
; VI-NEXT: v_or_b32_sdwa v17, v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v39
|
|
; VI-NEXT: v_or_b32_sdwa v11, v18, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v38
|
|
; VI-NEXT: v_or_b32_sdwa v19, v5, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v37
|
|
; VI-NEXT: v_or_b32_sdwa v20, v6, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v36
|
|
; VI-NEXT: v_or_b32_sdwa v21, v7, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v35
|
|
; VI-NEXT: v_or_b32_sdwa v22, v8, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v34
|
|
; VI-NEXT: v_or_b32_sdwa v23, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, v30
|
|
; VI-NEXT: v_mov_b32_e32 v1, v31
|
|
; VI-NEXT: v_mov_b32_e32 v2, v32
|
|
; VI-NEXT: v_mov_b32_e32 v3, v33
|
|
; VI-NEXT: v_mov_b32_e32 v4, v24
|
|
; VI-NEXT: v_mov_b32_e32 v5, v25
|
|
; VI-NEXT: v_mov_b32_e32 v6, v26
|
|
; VI-NEXT: v_mov_b32_e32 v7, v27
|
|
; VI-NEXT: v_mov_b32_e32 v8, v28
|
|
; VI-NEXT: v_mov_b32_e32 v9, v29
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB53_4:
|
|
; VI-NEXT: ; implicit-def: $vgpr30
|
|
; VI-NEXT: ; implicit-def: $vgpr31
|
|
; VI-NEXT: ; implicit-def: $vgpr32
|
|
; VI-NEXT: ; implicit-def: $vgpr33
|
|
; VI-NEXT: ; implicit-def: $vgpr55
|
|
; VI-NEXT: ; implicit-def: $vgpr25
|
|
; VI-NEXT: ; implicit-def: $vgpr26
|
|
; VI-NEXT: ; implicit-def: $vgpr27
|
|
; VI-NEXT: ; implicit-def: $vgpr28
|
|
; VI-NEXT: ; implicit-def: $vgpr29
|
|
; VI-NEXT: ; implicit-def: $vgpr10
|
|
; 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: s_branch .LBB53_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v12f64_to_v48f16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_mov_b32_e32 v23, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v19, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v20, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v21, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v22, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v17, s26
|
|
; GFX9-NEXT: v_mov_b32_e32 v18, s27
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_mov_b32_e32 v15, s28
|
|
; GFX9-NEXT: v_mov_b32_e32 v16, s29
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB53_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB53_3
|
|
; GFX9-NEXT: .LBB53_2: ; %cmp.true
|
|
; 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_add_f64 v[15:16], v[15:16], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
|
|
; GFX9-NEXT: v_add_f64 v[23:24], v[23:24], 1.0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v23
|
|
; GFX9-NEXT: .LBB53_3: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v30, v30, 16, v23
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v24
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v55, 16, v13
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v26, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v27, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v21
|
|
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v50, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v25, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v28, v28, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v22
|
|
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v49, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v3
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v29, v29, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v48, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v32, v32, 16, v19
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v10, 16, v11
|
|
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v39, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v33, v33, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v38, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v37, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v7
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v36, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v31, v31, 16, v23
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v35, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v54, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v53, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v52, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v34, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, v30
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, v31
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, v32
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, v25
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, v26
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, v27
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, v28
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, v29
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB53_4:
|
|
; GFX9-NEXT: ; implicit-def: $vgpr30
|
|
; GFX9-NEXT: ; implicit-def: $vgpr31
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: ; implicit-def: $vgpr33
|
|
; GFX9-NEXT: ; implicit-def: $vgpr55
|
|
; GFX9-NEXT: ; implicit-def: $vgpr25
|
|
; GFX9-NEXT: ; implicit-def: $vgpr26
|
|
; GFX9-NEXT: ; implicit-def: $vgpr27
|
|
; GFX9-NEXT: ; implicit-def: $vgpr28
|
|
; GFX9-NEXT: ; implicit-def: $vgpr29
|
|
; GFX9-NEXT: ; implicit-def: $vgpr10
|
|
; 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: s_branch .LBB53_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v12f64_to_v48f16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v23, v5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, v4 :: v_dual_mov_b32 v21, v3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v19, v1
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v5, s17
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v7, s19
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v9, s21
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v11, s23
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v13, s25
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v16, s28
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB53_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB53_3
|
|
; GFX11-TRUE16-NEXT: .LBB53_2: ; %cmp.true
|
|
; 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: v_lshrrev_b32_e32 v24, 16, v23
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; GFX11-TRUE16-NEXT: .LBB53_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v55, v55 :: v_dual_mov_b32 v54, v54
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v53 :: v_dual_mov_b32 v52, v52
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v51, v51 :: v_dual_mov_b32 v50, v50
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v49, v49 :: v_dual_mov_b32 v48, v48
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, v39 :: v_dual_mov_b32 v38, v38
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, v37 :: v_dual_mov_b32 v36, v36
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, v35 :: v_dual_mov_b32 v34, v34
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, v33 :: v_dual_mov_b32 v32, v32
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, v31 :: v_dual_mov_b32 v30, v30
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v29 :: v_dual_mov_b32 v28, v28
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v27 :: v_dual_mov_b32 v26, v26
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v25 :: v_dual_mov_b32 v24, v24
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v31.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v28.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v27.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v26.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v25.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v24.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB53_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB53_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v12f64_to_v48f16_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v25, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v23, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v21, s17
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v19, s21
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s24 :: v_dual_mov_b32 v10, s25
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s0, 0
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s1, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB53_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v13, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v6, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB53_3
|
|
; GFX11-FAKE16-NEXT: .LBB53_2: ; %cmp.true
|
|
; 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_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[9:10], v[9:10], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[18:19], v[18:19], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[20:21], v[20:21], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[22:23], v[22:23], 1.0
|
|
; GFX11-FAKE16-NEXT: v_add_f64 v[24:25], v[24:25], 1.0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v13, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v6, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: .LBB53_3: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v23, 0xffff, v23
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v29, v29, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v6, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v53, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v52, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v54, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v27, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v19, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v51, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v35, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v28, v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v20, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v13, 16, v20
|
|
; 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 v17, 0xffff, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v33, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v55, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v50, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v49, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v48, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v39, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v38, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v37, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v36, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v34, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v32, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v31, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v30, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v0, v24
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v28
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB53_4:
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr55
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr54
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr27
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr6
|
|
; 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: $vgpr13
|
|
; 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: s_branch .LBB53_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <12 x double> %a, splat (double 1.000000e+00)
|
|
%a2 = bitcast <12 x double> %a1 to <48 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <12 x double> %a to <48 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x half> %phi
|
|
}
|
|
|
|
define <12 x double> @bitcast_v48f16_to_v12f64(<48 x half> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v48f16_to_v12f64:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v14
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:24
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:20
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:32
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:28
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:40
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:36
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:48
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:44
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:56
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:52
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:64
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:60
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v1
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, v5
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v6
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v11
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, v12
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v15
|
|
; 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_cvt_f16_f32_e32 v0, v20
|
|
; 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_cvt_f16_f32_e32 v0, v23
|
|
; 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_cvt_f16_f32_e32 v0, v22
|
|
; 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_cvt_f16_f32_e32 v0, v25
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v24
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v27
|
|
; 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_cvt_f16_f32_e32 v0, v26
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v29
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v28
|
|
; 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_cvt_f16_f32_e32 v0, v30
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v31
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v32
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v40
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v40, v63
|
|
; 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_cvt_f16_f32_e32 v0, v41
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v62
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v42
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, v61
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, v43
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v60
|
|
; 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_cvt_f16_f32_e32 v0, v44
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v59
|
|
; 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_cvt_f16_f32_e32 v0, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v45, v58
|
|
; 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_cvt_f16_f32_e32 v0, v46
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, v57
|
|
; 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_cvt_f16_f32_e32 v0, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v56
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 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 v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v33
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v55
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v49
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v35
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v45
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v43
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v41
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: v_or_b32_e32 v0, v54, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v52, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v50, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v48, v3
|
|
; SI-NEXT: v_or_b32_e32 v4, v38, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v36, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v34, v6
|
|
; SI-NEXT: v_or_b32_e32 v21, v44, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v42, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v40, v23
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; 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: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: ; kill: killed $vgpr24
|
|
; 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: buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; 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:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; SI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
|
|
; SI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v20, v46, v20
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; 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: buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v55
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v53
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v52
|
|
; 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, v51
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v48
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v38
|
|
; 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, v49
|
|
; 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, v36
|
|
; 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, v34
|
|
; 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, v39
|
|
; 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 v21, v47
|
|
; 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 v22, v46
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; 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, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v44
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v41
|
|
; 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, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v40
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; 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, v33
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; SI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v12
|
|
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_or_b32_e32 v11, v13, v11
|
|
; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; 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: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; 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, v14, v12
|
|
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; 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: 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_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v15
|
|
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_or_b32_e32 v14, v16, v14
|
|
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v18
|
|
; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_or_b32_e32 v17, v19, v17
|
|
; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v45
|
|
; SI-NEXT: v_or_b32_e32 v20, v22, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v43
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v23, v25, v23
|
|
; 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 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v48f16_to_v12f64:
|
|
; 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: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_mov_b32_e32 v32, v23
|
|
; VI-NEXT: v_mov_b32_e32 v33, v22
|
|
; VI-NEXT: v_mov_b32_e32 v34, v21
|
|
; VI-NEXT: v_mov_b32_e32 v35, v20
|
|
; VI-NEXT: v_mov_b32_e32 v36, v19
|
|
; VI-NEXT: v_mov_b32_e32 v37, v18
|
|
; VI-NEXT: v_mov_b32_e32 v38, v17
|
|
; VI-NEXT: v_mov_b32_e32 v39, v16
|
|
; VI-NEXT: v_mov_b32_e32 v48, v15
|
|
; VI-NEXT: v_mov_b32_e32 v49, v14
|
|
; VI-NEXT: v_mov_b32_e32 v50, v13
|
|
; VI-NEXT: v_mov_b32_e32 v51, v12
|
|
; VI-NEXT: v_mov_b32_e32 v52, v11
|
|
; VI-NEXT: v_mov_b32_e32 v53, v10
|
|
; VI-NEXT: v_mov_b32_e32 v54, v9
|
|
; VI-NEXT: v_mov_b32_e32 v55, v8
|
|
; VI-NEXT: v_mov_b32_e32 v40, v7
|
|
; VI-NEXT: v_mov_b32_e32 v41, v6
|
|
; VI-NEXT: v_mov_b32_e32 v42, v5
|
|
; VI-NEXT: v_mov_b32_e32 v43, v4
|
|
; VI-NEXT: v_mov_b32_e32 v44, v3
|
|
; VI-NEXT: v_mov_b32_e32 v45, v2
|
|
; VI-NEXT: v_mov_b32_e32 v46, v1
|
|
; VI-NEXT: v_mov_b32_e32 v47, v0
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; 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 v23, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v23, v47 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v23, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v2, v23, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v3, v23, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v4, v23, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v5, v23, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v6, v23, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v7, v23, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v8, v23, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v9, v23, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v10, v23, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v11, v23, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v12, v23, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v13, v23, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v14, v23, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v15, v23, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v16, v23, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v17, v23, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v18, v23, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v19, v23, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v20, v23, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v21, v23, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v22, v23, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v23, v23, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v46, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v45, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v44, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v4, v43, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v5, v42, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v6, v41, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v7, v40, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v8, v55, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v10, v53, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v52, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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 v23, 0x200
|
|
; VI-NEXT: v_add_f16_sdwa v0, v47, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v47
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_add_f16_sdwa v1, v46, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v2, 0x200, v46
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; VI-NEXT: v_add_f16_sdwa v2, v45, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v45
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_add_f16_sdwa v3, v44, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v44
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_add_f16_sdwa v4, v43, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v43
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_add_f16_sdwa v5, v42, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v42
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_add_f16_sdwa v6, v41, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v41
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_add_f16_sdwa v7, v40, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v40
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_add_f16_sdwa v8, v55, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v55
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_add_f16_sdwa v9, v54, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, 0x200, v54
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, 0x200, v53
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_add_f16_sdwa v11, v52, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, 0x200, v52
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_add_f16_sdwa v12, v51, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, 0x200, v51
|
|
; VI-NEXT: v_or_b32_e32 v12, v13, v12
|
|
; VI-NEXT: v_add_f16_sdwa v13, v50, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v50
|
|
; VI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; VI-NEXT: v_add_f16_sdwa v14, v49, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v48, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v39, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v38, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v37, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v36, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v35, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v34, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB54_4: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; 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: s_waitcnt vmcnt(0)
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48f16_to_v12f64:
|
|
; 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 v33, v22
|
|
; GFX9-NEXT: v_mov_b32_e32 v47, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v23
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v21
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v20
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v34
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v19
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v35
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v18
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v36
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v17
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v37
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v16
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v38
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v15
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v39
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v14
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v48
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, v13
|
|
; 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, v49
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, v12
|
|
; 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, v50
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, v11
|
|
; 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, v51
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, v10
|
|
; 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, v52
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v40, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v41, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v42, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v43, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v44, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v45, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v46, v1
|
|
; 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, v53
|
|
; 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, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v63, 16, v55
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v56, 16, v40
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v57, 16, v41
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v58, 16, v42
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v59, 16, v43
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v60, 16, v44
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v61, 16, v45
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v62, 16, v46
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v47
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; GFX9-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 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:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v54
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v53
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v52
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v51
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v50
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v49
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v33
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; 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: $vgpr63
|
|
; GFX9-NEXT: ; implicit-def: $vgpr56
|
|
; GFX9-NEXT: ; implicit-def: $vgpr57
|
|
; GFX9-NEXT: ; implicit-def: $vgpr58
|
|
; GFX9-NEXT: ; implicit-def: $vgpr59
|
|
; GFX9-NEXT: ; implicit-def: $vgpr60
|
|
; GFX9-NEXT: ; implicit-def: $vgpr61
|
|
; GFX9-NEXT: ; implicit-def: $vgpr62
|
|
; GFX9-NEXT: ; implicit-def: $vgpr32
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; GFX9-NEXT: ; implicit-def: $vgpr24
|
|
; GFX9-NEXT: ; kill: killed $vgpr24
|
|
; 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 v9, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_mov_b32 s6, 0x5040100
|
|
; GFX9-NEXT: v_perm_b32 v0, v32, v47, s6
|
|
; GFX9-NEXT: s_movk_i32 s7, 0x200
|
|
; GFX9-NEXT: v_perm_b32 v1, v62, v46, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v61, v45, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v60, v44, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v59, v43, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v58, v42, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v57, v41, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v56, v40, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v63, v55, 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(15)
|
|
; GFX9-NEXT: v_perm_b32 v9, v9, v54, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(14)
|
|
; GFX9-NEXT: v_perm_b32 v10, v10, v53, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(13)
|
|
; GFX9-NEXT: v_perm_b32 v11, v11, v52, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(12)
|
|
; GFX9-NEXT: v_perm_b32 v12, v12, v51, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(11)
|
|
; GFX9-NEXT: v_perm_b32 v13, v13, v50, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(10)
|
|
; GFX9-NEXT: v_perm_b32 v14, v14, v49, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(9)
|
|
; GFX9-NEXT: v_perm_b32 v15, v15, v48, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(8)
|
|
; GFX9-NEXT: v_perm_b32 v16, v16, v39, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(7)
|
|
; GFX9-NEXT: v_perm_b32 v17, v17, v38, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(6)
|
|
; GFX9-NEXT: v_perm_b32 v18, v18, v37, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(5)
|
|
; GFX9-NEXT: v_perm_b32 v19, v19, v36, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(4)
|
|
; GFX9-NEXT: v_perm_b32 v20, v20, v35, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(3)
|
|
; GFX9-NEXT: v_perm_b32 v21, v21, v34, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(2)
|
|
; GFX9-NEXT: v_perm_b32 v22, v22, v33, s6
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: v_perm_b32 v23, v24, v23, 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: .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_v48f16_to_v12f64:
|
|
; 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, v24
|
|
; 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: .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_v48f16_to_v12f64:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v24
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v53, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v54, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v55, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v64, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v52, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v51, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v50, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v49, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v48, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v39, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v38, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v37, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v36, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v35, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v34, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v33, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v32, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v31, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v30, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v29, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v28, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v27, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v26, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v25, v23, 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: .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 <48 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <48 x half> %a1 to <12 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x half> %a to <12 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x double> %phi
|
|
}
|
|
|
|
define inreg <12 x double> @bitcast_v48f16_to_v12f64_scalar(<48 x half> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v48f16_to_v12f64_scalar:
|
|
; 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:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v62, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v0
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v63, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v61, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v60, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v59, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v58, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v57, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v56, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v45, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v0, s17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, s16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, s19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, s18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, s21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, s20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, s23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, s22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, s25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, s24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, s27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v30, s28
|
|
; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v31
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v34
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v35
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v35, s26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, s29
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v40
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: s_cbranch_scc0 .LBB55_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v0, v8, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v7, v1
|
|
; SI-NEXT: v_or_b32_e32 v2, v6, v2
|
|
; SI-NEXT: v_or_b32_e32 v3, v5, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v34
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v62
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v52
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v48
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v36
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v60
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v58
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v46
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v42
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v27
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v4, v54, v4
|
|
; SI-NEXT: v_or_b32_e32 v5, v35, v5
|
|
; SI-NEXT: v_or_b32_e32 v6, v30, v6
|
|
; SI-NEXT: v_or_b32_e32 v7, v55, v7
|
|
; SI-NEXT: v_or_b32_e32 v8, v63, v8
|
|
; SI-NEXT: v_or_b32_e32 v9, v43, v9
|
|
; SI-NEXT: v_or_b32_e32 v10, v49, v10
|
|
; SI-NEXT: v_or_b32_e32 v11, v39, v11
|
|
; SI-NEXT: v_or_b32_e32 v12, v37, v12
|
|
; SI-NEXT: v_or_b32_e32 v13, v61, v13
|
|
; SI-NEXT: v_or_b32_e32 v14, v33, v14
|
|
; SI-NEXT: v_or_b32_e32 v15, v59, v15
|
|
; SI-NEXT: v_or_b32_e32 v16, v57, v16
|
|
; SI-NEXT: v_or_b32_e32 v17, v47, v17
|
|
; SI-NEXT: v_or_b32_e32 v18, v45, v18
|
|
; SI-NEXT: v_or_b32_e32 v19, v25, v19
|
|
; SI-NEXT: v_or_b32_e32 v20, v41, v20
|
|
; SI-NEXT: v_or_b32_e32 v21, v28, v21
|
|
; SI-NEXT: v_or_b32_e32 v22, v26, v22
|
|
; SI-NEXT: v_or_b32_e32 v23, v31, v23
|
|
; SI-NEXT: s_cbranch_execnz .LBB55_3
|
|
; SI-NEXT: .LBB55_2: ; %cmp.true
|
|
; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v35
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v30
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v55
|
|
; 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_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; 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 v9, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v43
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v49
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v48
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v39
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v37
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v60
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v33
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; 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_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v59
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v56
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v47
|
|
; 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, v45
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v42
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v41
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v28
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v0, 0x38000000, 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 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: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_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 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; SI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; 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, v51
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; 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, v34
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; 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, v62
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; 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, v53
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; 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, v52
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; 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, v50
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; SI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v38
|
|
; SI-NEXT: v_or_b32_e32 v11, v13, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v36
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v12, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v61
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_or_b32_e32 v13, v14, v13
|
|
; SI-NEXT: v_lshlrev_b32_e32 v14, 16, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v32
|
|
; SI-NEXT: v_or_b32_e32 v14, v16, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v58
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v15, v17, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v57
|
|
; SI-NEXT: v_lshlrev_b32_e32 v16, 16, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v46
|
|
; SI-NEXT: v_or_b32_e32 v17, v19, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v44
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v18, v20, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v31
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v20, 16, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v29
|
|
; SI-NEXT: v_or_b32_e32 v20, v22, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v27
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
|
|
; SI-NEXT: v_or_b32_e32 v21, v23, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v23, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v22
|
|
; SI-NEXT: v_add_f32_e32 v23, 0x38000000, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
|
|
; SI-NEXT: v_or_b32_e32 v22, v23, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v23, v25, v23
|
|
; SI-NEXT: .LBB55_3: ; %end
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB55_4:
|
|
; SI-NEXT: v_mov_b32_e32 v40, v31
|
|
; SI-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v54, v33
|
|
; SI-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v51, v32
|
|
; SI-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v34, v59
|
|
; SI-NEXT: v_mov_b32_e32 v59, v46
|
|
; SI-NEXT: v_mov_b32_e32 v46, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v24
|
|
; SI-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v35, v60
|
|
; SI-NEXT: v_mov_b32_e32 v60, v47
|
|
; SI-NEXT: v_mov_b32_e32 v47, v42
|
|
; SI-NEXT: v_mov_b32_e32 v42, v26
|
|
; SI-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_mov_b32_e32 v55, v52
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: v_mov_b32_e32 v53, v43
|
|
; SI-NEXT: v_mov_b32_e32 v52, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v61
|
|
; SI-NEXT: v_mov_b32_e32 v61, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v25
|
|
; SI-NEXT: v_mov_b32_e32 v43, v27
|
|
; SI-NEXT: v_mov_b32_e32 v33, v62
|
|
; SI-NEXT: v_mov_b32_e32 v62, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v44
|
|
; SI-NEXT: v_mov_b32_e32 v44, v28
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_mov_b32_e32 v63, v58
|
|
; SI-NEXT: v_mov_b32_e32 v58, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v29
|
|
; SI-NEXT: v_mov_b32_e32 v32, v30
|
|
; 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: v_mov_b32_e32 v29, v45
|
|
; SI-NEXT: v_mov_b32_e32 v45, v58
|
|
; SI-NEXT: v_mov_b32_e32 v58, v63
|
|
; SI-NEXT: v_mov_b32_e32 v27, v43
|
|
; SI-NEXT: v_mov_b32_e32 v25, v56
|
|
; SI-NEXT: v_mov_b32_e32 v56, v61
|
|
; SI-NEXT: v_mov_b32_e32 v61, v36
|
|
; SI-NEXT: v_mov_b32_e32 v36, v37
|
|
; SI-NEXT: v_mov_b32_e32 v37, v38
|
|
; SI-NEXT: v_mov_b32_e32 v38, v39
|
|
; SI-NEXT: v_mov_b32_e32 v39, v48
|
|
; SI-NEXT: v_mov_b32_e32 v48, v49
|
|
; SI-NEXT: v_mov_b32_e32 v49, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v52
|
|
; SI-NEXT: v_mov_b32_e32 v43, v53
|
|
; SI-NEXT: v_mov_b32_e32 v52, v55
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v30, v32
|
|
; SI-NEXT: v_mov_b32_e32 v28, v44
|
|
; SI-NEXT: v_mov_b32_e32 v44, v57
|
|
; SI-NEXT: v_mov_b32_e32 v57, v62
|
|
; SI-NEXT: v_mov_b32_e32 v62, v33
|
|
; SI-NEXT: v_mov_b32_e32 v26, v42
|
|
; SI-NEXT: v_mov_b32_e32 v42, v47
|
|
; SI-NEXT: v_mov_b32_e32 v47, v60
|
|
; SI-NEXT: v_mov_b32_e32 v60, v35
|
|
; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v24, v41
|
|
; SI-NEXT: v_mov_b32_e32 v41, v46
|
|
; SI-NEXT: v_mov_b32_e32 v46, v59
|
|
; SI-NEXT: v_mov_b32_e32 v59, v34
|
|
; SI-NEXT: v_mov_b32_e32 v32, v51
|
|
; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v33, v54
|
|
; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_mov_b32_e32 v31, v40
|
|
; SI-NEXT: s_branch .LBB55_2
|
|
;
|
|
; VI-LABEL: bitcast_v48f16_to_v12f64_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_mov_b32_e32 v32, v9
|
|
; VI-NEXT: v_mov_b32_e32 v33, v8
|
|
; VI-NEXT: v_mov_b32_e32 v34, v7
|
|
; VI-NEXT: v_mov_b32_e32 v35, v6
|
|
; VI-NEXT: v_mov_b32_e32 v36, v5
|
|
; VI-NEXT: v_mov_b32_e32 v37, v4
|
|
; VI-NEXT: v_mov_b32_e32 v38, v3
|
|
; VI-NEXT: v_mov_b32_e32 v39, v2
|
|
; VI-NEXT: v_mov_b32_e32 v48, v1
|
|
; VI-NEXT: v_mov_b32_e32 v49, v0
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: s_cbranch_scc0 .LBB55_4
|
|
; VI-NEXT: ; %bb.1: ; %cmp.false
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s43, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s44, s42, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s44
|
|
; VI-NEXT: s_and_b32 s44, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s45, s41, 16
|
|
; VI-NEXT: s_or_b32 s44, s44, s45
|
|
; VI-NEXT: s_and_b32 s45, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s46, s40, 16
|
|
; VI-NEXT: s_or_b32 s45, s45, s46
|
|
; VI-NEXT: s_and_b32 s46, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s47, s15, 16
|
|
; VI-NEXT: s_or_b32 s46, s46, s47
|
|
; VI-NEXT: s_and_b32 s47, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s56, s14, 16
|
|
; VI-NEXT: v_mov_b32_e32 v0, 16
|
|
; VI-NEXT: s_or_b32 s47, s47, s56
|
|
; VI-NEXT: s_and_b32 s56, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s57, s13, 16
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s56, s56, s57
|
|
; VI-NEXT: s_and_b32 s57, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s58, s12, 16
|
|
; VI-NEXT: v_or_b32_sdwa v14, v49, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s57, s57, s58
|
|
; VI-NEXT: s_and_b32 s58, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s59, s11, 16
|
|
; VI-NEXT: v_or_b32_sdwa v15, v48, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s58, s58, s59
|
|
; VI-NEXT: s_and_b32 s59, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s60, s10, 16
|
|
; VI-NEXT: v_or_b32_sdwa v16, v39, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s59, s59, s60
|
|
; VI-NEXT: s_and_b32 s60, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s61, s9, 16
|
|
; VI-NEXT: v_or_b32_sdwa v17, v38, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s60, s60, s61
|
|
; VI-NEXT: s_and_b32 s61, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s62, s8, 16
|
|
; VI-NEXT: v_or_b32_sdwa v18, v37, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s61, s61, s62
|
|
; VI-NEXT: s_and_b32 s62, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s63, s7, 16
|
|
; VI-NEXT: v_or_b32_sdwa v19, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s62, s62, s63
|
|
; VI-NEXT: s_and_b32 s63, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s72, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v20, v35, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: s_or_b32 s63, s63, s72
|
|
; VI-NEXT: v_or_b32_sdwa v21, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v1, v0, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_lshlrev_b32_sdwa v0, v0, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; VI-NEXT: v_or_b32_sdwa v22, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-NEXT: v_mov_b32_e32 v2, s44
|
|
; VI-NEXT: v_mov_b32_e32 v3, s45
|
|
; VI-NEXT: v_mov_b32_e32 v4, s46
|
|
; VI-NEXT: v_mov_b32_e32 v5, s47
|
|
; VI-NEXT: v_mov_b32_e32 v6, s56
|
|
; VI-NEXT: v_mov_b32_e32 v7, s57
|
|
; VI-NEXT: v_mov_b32_e32 v8, s58
|
|
; VI-NEXT: v_mov_b32_e32 v9, s59
|
|
; VI-NEXT: v_mov_b32_e32 v10, s60
|
|
; VI-NEXT: v_mov_b32_e32 v11, s61
|
|
; VI-NEXT: v_mov_b32_e32 v12, s62
|
|
; VI-NEXT: v_mov_b32_e32 v13, s63
|
|
; VI-NEXT: s_cbranch_execnz .LBB55_3
|
|
; VI-NEXT: .LBB55_2: ; %cmp.true
|
|
; VI-NEXT: v_mov_b32_e32 v13, 0x200
|
|
; VI-NEXT: v_mov_b32_e32 v0, s43
|
|
; VI-NEXT: v_add_f16_sdwa v0, v0, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v1, s16, v13
|
|
; VI-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; VI-NEXT: v_mov_b32_e32 v1, s42
|
|
; VI-NEXT: v_add_f16_sdwa v1, v1, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v2, s17, v13
|
|
; VI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; VI-NEXT: v_mov_b32_e32 v2, s41
|
|
; VI-NEXT: v_add_f16_sdwa v2, v2, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v3, s18, v13
|
|
; VI-NEXT: v_or_b32_e32 v2, v3, v2
|
|
; VI-NEXT: v_mov_b32_e32 v3, s40
|
|
; VI-NEXT: v_add_f16_sdwa v3, v3, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v4, s19, v13
|
|
; VI-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; VI-NEXT: v_mov_b32_e32 v4, s15
|
|
; VI-NEXT: v_add_f16_sdwa v4, v4, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v5, s20, v13
|
|
; VI-NEXT: v_or_b32_e32 v4, v5, v4
|
|
; VI-NEXT: v_mov_b32_e32 v5, s14
|
|
; VI-NEXT: v_add_f16_sdwa v5, v5, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v6, s21, v13
|
|
; VI-NEXT: v_or_b32_e32 v5, v6, v5
|
|
; VI-NEXT: v_mov_b32_e32 v6, s13
|
|
; VI-NEXT: v_add_f16_sdwa v6, v6, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v7, s22, v13
|
|
; VI-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; VI-NEXT: v_mov_b32_e32 v7, s12
|
|
; VI-NEXT: v_add_f16_sdwa v7, v7, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v8, s23, v13
|
|
; VI-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; VI-NEXT: v_mov_b32_e32 v8, s11
|
|
; VI-NEXT: v_add_f16_sdwa v8, v8, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v9, s24, v13
|
|
; VI-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; VI-NEXT: v_mov_b32_e32 v9, s10
|
|
; VI-NEXT: v_add_f16_sdwa v9, v9, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v10, s25, v13
|
|
; VI-NEXT: v_or_b32_e32 v9, v10, v9
|
|
; VI-NEXT: v_mov_b32_e32 v10, s9
|
|
; VI-NEXT: v_add_f16_sdwa v10, v10, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v11, s26, v13
|
|
; VI-NEXT: v_or_b32_e32 v10, v11, v10
|
|
; VI-NEXT: v_mov_b32_e32 v11, s8
|
|
; VI-NEXT: v_add_f16_sdwa v11, v11, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v12, s27, v13
|
|
; VI-NEXT: v_or_b32_e32 v11, v12, v11
|
|
; VI-NEXT: v_mov_b32_e32 v12, s7
|
|
; VI-NEXT: v_add_f16_sdwa v12, v12, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v14, s28, v13
|
|
; VI-NEXT: v_or_b32_e32 v12, v14, v12
|
|
; VI-NEXT: v_mov_b32_e32 v14, s6
|
|
; VI-NEXT: v_add_f16_sdwa v14, v14, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v13, s29, v13
|
|
; VI-NEXT: v_mov_b32_e32 v23, 0x200
|
|
; VI-NEXT: v_or_b32_e32 v13, v13, v14
|
|
; VI-NEXT: v_add_f16_sdwa v14, v49, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v49
|
|
; VI-NEXT: v_or_b32_e32 v14, v15, v14
|
|
; VI-NEXT: v_add_f16_sdwa v15, v48, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v48
|
|
; VI-NEXT: v_or_b32_e32 v15, v16, v15
|
|
; VI-NEXT: v_add_f16_sdwa v16, v39, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v39
|
|
; VI-NEXT: v_or_b32_e32 v16, v17, v16
|
|
; VI-NEXT: v_add_f16_sdwa v17, v38, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v38
|
|
; VI-NEXT: v_or_b32_e32 v17, v18, v17
|
|
; VI-NEXT: v_add_f16_sdwa v18, v37, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v37
|
|
; VI-NEXT: v_or_b32_e32 v18, v19, v18
|
|
; VI-NEXT: v_add_f16_sdwa v19, v36, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v36
|
|
; VI-NEXT: v_or_b32_e32 v19, v20, v19
|
|
; VI-NEXT: v_add_f16_sdwa v20, v35, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v35
|
|
; VI-NEXT: v_or_b32_e32 v20, v21, v20
|
|
; VI-NEXT: v_add_f16_sdwa v21, v34, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v34
|
|
; VI-NEXT: v_or_b32_e32 v21, v22, v21
|
|
; VI-NEXT: v_add_f16_sdwa v22, v33, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v33
|
|
; VI-NEXT: v_or_b32_e32 v22, v24, v22
|
|
; VI-NEXT: v_add_f16_sdwa v23, v32, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v32
|
|
; VI-NEXT: v_or_b32_e32 v23, v24, v23
|
|
; VI-NEXT: .LBB55_3: ; %end
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB55_4:
|
|
; 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_branch .LBB55_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v48f16_to_v12f64_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, v9
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, v8
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, v7
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, v6
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, v5
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, v4
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, v3
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, v2
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, v1
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, v0
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: s_lshr_b32 s40, s29, 16
|
|
; GFX9-NEXT: s_lshr_b32 s41, s28, 16
|
|
; GFX9-NEXT: s_lshr_b32 s42, s27, 16
|
|
; GFX9-NEXT: s_lshr_b32 s43, 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: buffer_store_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s32 ; 4-byte Folded Spill
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v32
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v33
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v34
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v35
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v36
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v37
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v40, 16, v38
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v41, 16, v39
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v42, 16, v48
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v43, 16, v49
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s17, s7
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s8, s18, s8
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s9, s19, s9
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s10, s20, s10
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s11, s21, s11
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s12, s22, s12
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s13, s23, s13
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s14, s24, s14
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s15, s25, s15
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s16, s26, s43
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s17, s27, s42
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s18, s28, s41
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s19, s29, s40
|
|
; GFX9-NEXT: s_cbranch_scc0 .LBB55_4
|
|
; GFX9-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v49
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v48
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v39
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v38
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v37
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v36
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v35
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v34
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v33
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, s6
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v10, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v11, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v12, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, s19
|
|
; GFX9-NEXT: s_cbranch_execnz .LBB55_3
|
|
; GFX9-NEXT: .LBB55_2: ; %cmp.true
|
|
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff, v49
|
|
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff, v48
|
|
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff, v39
|
|
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff, v38
|
|
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff, v37
|
|
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff, v36
|
|
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff, v35
|
|
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff, v34
|
|
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff, v33
|
|
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff, v32
|
|
; GFX9-NEXT: s_movk_i32 s4, 0x200
|
|
; GFX9-NEXT: v_mov_b32_e32 v13, 0x200
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v43, 16, v14
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v42, 16, v15
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v41, 16, v16
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v40, 16, v17
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v55, 16, v18
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v54, 16, v19
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v53, 16, v20
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v52, 16, v21
|
|
; GFX9-NEXT: v_lshl_or_b32 v22, v51, 16, v22
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v50, 16, v23
|
|
; GFX9-NEXT: v_pk_add_f16 v0, s6, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, s7, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, s8, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, s9, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, s10, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, s11, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, s12, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, s13, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, s14, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v9, s15, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v10, s16, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v11, s17, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v12, s18, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v13, s19, v13 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v14, v14, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v15, v15, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v16, v16, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v17, v17, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v18, v18, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v19, v19, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v20, v20, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v21, v21, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v22, v22, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v23, v23, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: .LBB55_3: ; %end
|
|
; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s32 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX9-NEXT: .LBB55_4:
|
|
; 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_branch .LBB55_2
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v12f64_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB55_4
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; 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, s4 :: v_dual_mov_b32 v5, s5
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB55_3
|
|
; GFX11-TRUE16-NEXT: .LBB55_2: ; %cmp.true
|
|
; 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, s4 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, s17 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: .LBB55_3: ; %end
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-TRUE16-NEXT: .LBB55_4:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB55_2
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v48f16_to_v12f64_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v51, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v50, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v49, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v38, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s41, s29, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s42, s28, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s15, s27, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s43, s26, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s14, s25, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s13, s24, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s12, s23, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s11, s22, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s10, s21, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s9, s20, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s19, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s7, s18, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s17, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s44, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s45, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s46, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s40, 0
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s4, s0, s4
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s1, s1, s46
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s45
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s44
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s5, s16, s5
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s6, s17, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s7, s18, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s8, s19, s8
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s9, s20, s9
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s10, s21, s10
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s11, s22, s11
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s12, s23, s12
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s13, s24, s13
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s14, s25, s14
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s0, s26, s43
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-FAKE16-NEXT: s_cbranch_scc0 .LBB55_4
|
|
; GFX11-FAKE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
|
|
; GFX11-FAKE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB55_3
|
|
; GFX11-FAKE16-NEXT: .LBB55_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v37, 16, v51
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v36, 16, v50
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v35, 16, v49
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v33, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v32, 16, v38
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v0, 0x200, s4 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 v2, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, s17 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: .LBB55_3: ; %end
|
|
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
|
|
; GFX11-FAKE16-NEXT: .LBB55_4:
|
|
; GFX11-FAKE16-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
|
|
; GFX11-FAKE16-NEXT: s_branch .LBB55_2
|
|
%cmp = icmp eq i32 %b, 0
|
|
br i1 %cmp, label %cmp.true, label %cmp.false
|
|
|
|
cmp.true:
|
|
%a1 = fadd <48 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <48 x half> %a1 to <12 x double>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x half> %a to <12 x double>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <12 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <12 x double> %phi
|
|
}
|
|
|
|
define <48 x half> @bitcast_v48i16_to_v48f16(<48 x i16> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v48i16_to_v48f16:
|
|
; 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:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:68
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:64
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:60
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:56
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:52
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:48
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:44
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:40
|
|
; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:72
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:36
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:32
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:28
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:24
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:20
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $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 .LBB56_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v4
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v7
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v10
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, v14
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v40
|
|
; SI-NEXT: ; implicit-def: $vgpr2
|
|
; SI-NEXT: ; implicit-def: $vgpr3
|
|
; SI-NEXT: ; implicit-def: $vgpr4
|
|
; SI-NEXT: ; implicit-def: $vgpr5
|
|
; SI-NEXT: ; implicit-def: $vgpr6
|
|
; 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: $vgpr13
|
|
; SI-NEXT: ; implicit-def: $vgpr14
|
|
; SI-NEXT: ; implicit-def: $vgpr15
|
|
; SI-NEXT: ; implicit-def: $vgpr16
|
|
; SI-NEXT: ; implicit-def: $vgpr17
|
|
; SI-NEXT: ; implicit-def: $vgpr18
|
|
; SI-NEXT: ; implicit-def: $vgpr19
|
|
; SI-NEXT: ; implicit-def: $vgpr40
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v20
|
|
; SI-NEXT: ; implicit-def: $vgpr20
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v21
|
|
; SI-NEXT: ; implicit-def: $vgpr21
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v22
|
|
; SI-NEXT: ; implicit-def: $vgpr22
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v23
|
|
; SI-NEXT: ; implicit-def: $vgpr23
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v24
|
|
; SI-NEXT: ; implicit-def: $vgpr24
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v25
|
|
; SI-NEXT: ; implicit-def: $vgpr25
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v26
|
|
; SI-NEXT: ; implicit-def: $vgpr26
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v27
|
|
; SI-NEXT: ; implicit-def: $vgpr27
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v28
|
|
; SI-NEXT: ; implicit-def: $vgpr28
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v29
|
|
; SI-NEXT: ; implicit-def: $vgpr29
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v30
|
|
; SI-NEXT: ; implicit-def: $vgpr30
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v54
|
|
; SI-NEXT: ; implicit-def: $vgpr54
|
|
; 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_cvt_f32_f16_e32 v1, v55
|
|
; SI-NEXT: ; implicit-def: $vgpr55
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v56
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; 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_cvt_f32_f16_e32 v1, v57
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v58
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v59
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; 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_cvt_f32_f16_e32 v1, v60
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; 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_cvt_f32_f16_e32 v1, v61
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v62
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; 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_cvt_f32_f16_e32 v1, v63
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; 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_cvt_f32_f16_e32 v1, v41
|
|
; SI-NEXT: ; implicit-def: $vgpr41
|
|
; 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_cvt_f32_f16_e32 v1, v42
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; 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_cvt_f32_f16_e32 v1, v43
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v44
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v45
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; 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_cvt_f32_f16_e32 v1, v46
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v47
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: ; implicit-def: $vgpr1
|
|
; 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 v13, vcc, 3, v13
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v13
|
|
; SI-NEXT: v_add_i32_e32 v15, vcc, 3, v15
|
|
; SI-NEXT: v_add_i32_e32 v17, vcc, 3, v17
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v15
|
|
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v18
|
|
; SI-NEXT: v_add_i32_e32 v19, vcc, 3, v19
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v17
|
|
; SI-NEXT: v_add_i32_e32 v20, vcc, 3, v20
|
|
; SI-NEXT: v_add_i32_e32 v21, vcc, 3, v21
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v18
|
|
; SI-NEXT: v_add_i32_e32 v22, vcc, 3, v22
|
|
; SI-NEXT: v_add_i32_e32 v23, vcc, 3, v23
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v19
|
|
; SI-NEXT: v_add_i32_e32 v24, vcc, 3, v24
|
|
; SI-NEXT: v_add_i32_e32 v25, vcc, 3, v25
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v20
|
|
; SI-NEXT: v_add_i32_e32 v26, vcc, 3, v26
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 3, v27
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v21
|
|
; SI-NEXT: v_add_i32_e32 v28, vcc, 3, v28
|
|
; SI-NEXT: v_add_i32_e32 v29, vcc, 3, v29
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v22
|
|
; SI-NEXT: v_add_i32_e32 v30, vcc, 3, v30
|
|
; SI-NEXT: v_add_i32_e32 v54, vcc, 3, v54
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v23
|
|
; SI-NEXT: v_add_i32_e32 v55, vcc, 3, v55
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_add_i32_e32 v56, vcc, 3, v56
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v24
|
|
; SI-NEXT: v_add_i32_e32 v57, vcc, 3, v57
|
|
; SI-NEXT: v_add_i32_e32 v58, vcc, 3, v58
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v25
|
|
; SI-NEXT: v_add_i32_e32 v59, vcc, 3, v59
|
|
; SI-NEXT: v_add_i32_e32 v60, vcc, 3, v60
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v26
|
|
; SI-NEXT: v_add_i32_e32 v61, vcc, 3, v61
|
|
; SI-NEXT: v_add_i32_e32 v62, vcc, 3, v62
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v27
|
|
; SI-NEXT: v_add_i32_e32 v63, vcc, 3, v63
|
|
; SI-NEXT: v_add_i32_e32 v41, vcc, 3, v41
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v28
|
|
; SI-NEXT: v_add_i32_e32 v42, vcc, 3, v42
|
|
; SI-NEXT: v_add_i32_e32 v43, vcc, 3, v43
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v29
|
|
; SI-NEXT: v_add_i32_e32 v44, vcc, 3, v44
|
|
; SI-NEXT: v_add_i32_e32 v45, vcc, 3, v45
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v30
|
|
; SI-NEXT: v_add_i32_e32 v46, vcc, 3, v46
|
|
; SI-NEXT: v_add_i32_e32 v47, vcc, 3, v47
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v54
|
|
; SI-NEXT: v_add_i32_e32 v40, vcc, 3, v40
|
|
; SI-NEXT: v_add_i32_e32 v16, vcc, 3, v16
|
|
; 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_cvt_f32_f16_e32 v1, v55
|
|
; SI-NEXT: v_add_i32_e32 v14, vcc, 3, v14
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v56
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; 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_cvt_f32_f16_e32 v1, v57
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v58
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v59
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; 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_cvt_f32_f16_e32 v1, v60
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; 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_cvt_f32_f16_e32 v1, v61
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v4
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v62
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, v7
|
|
; 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_cvt_f32_f16_e32 v1, v63
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v10
|
|
; 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_cvt_f32_f16_e32 v1, v41
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, v14
|
|
; 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_cvt_f32_f16_e32 v1, v42
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v40
|
|
; 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_cvt_f32_f16_e32 v1, v43
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v44
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v45
|
|
; 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_cvt_f32_f16_e32 v1, v46
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v47
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 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: v_cvt_f16_f32_e32 v1, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v49
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v53
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v3
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
|
|
; SI-NEXT: v_or_b32_e32 v2, v4, v2
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v51
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v38
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v48
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v36
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v39
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v34
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v37
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v32
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v35
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v33
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v31
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v48i16_to_v48f16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v5
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; 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 v55, 3, v55
|
|
; VI-NEXT: v_add_u16_e32 v1, 3, v1
|
|
; VI-NEXT: v_add_u16_e32 v54, 3, v54
|
|
; VI-NEXT: v_add_u16_e32 v2, 3, v2
|
|
; VI-NEXT: v_add_u16_e32 v53, 3, v53
|
|
; VI-NEXT: v_add_u16_e32 v3, 3, v3
|
|
; VI-NEXT: v_add_u16_e32 v52, 3, v52
|
|
; VI-NEXT: v_add_u16_e32 v4, 3, v4
|
|
; VI-NEXT: v_add_u16_e32 v24, 3, v24
|
|
; VI-NEXT: v_add_u16_e32 v5, 3, v5
|
|
; VI-NEXT: v_add_u16_e32 v51, 3, v51
|
|
; VI-NEXT: v_add_u16_e32 v6, 3, v6
|
|
; VI-NEXT: v_add_u16_e32 v50, 3, v50
|
|
; VI-NEXT: v_add_u16_e32 v7, 3, v7
|
|
; VI-NEXT: v_add_u16_e32 v49, 3, v49
|
|
; VI-NEXT: v_add_u16_e32 v8, 3, v8
|
|
; VI-NEXT: v_add_u16_e32 v48, 3, v48
|
|
; VI-NEXT: v_add_u16_e32 v9, 3, v9
|
|
; VI-NEXT: v_add_u16_e32 v39, 3, v39
|
|
; VI-NEXT: v_add_u16_e32 v10, 3, v10
|
|
; VI-NEXT: v_add_u16_e32 v38, 3, v38
|
|
; VI-NEXT: v_add_u16_e32 v11, 3, v11
|
|
; VI-NEXT: v_add_u16_e32 v37, 3, v37
|
|
; VI-NEXT: v_add_u16_e32 v12, 3, v12
|
|
; VI-NEXT: v_add_u16_e32 v36, 3, v36
|
|
; VI-NEXT: v_add_u16_e32 v13, 3, v13
|
|
; VI-NEXT: v_add_u16_e32 v35, 3, v35
|
|
; VI-NEXT: v_add_u16_e32 v14, 3, v14
|
|
; VI-NEXT: v_add_u16_e32 v34, 3, v34
|
|
; VI-NEXT: v_add_u16_e32 v15, 3, v15
|
|
; VI-NEXT: v_add_u16_e32 v33, 3, v33
|
|
; VI-NEXT: v_add_u16_e32 v16, 3, v16
|
|
; VI-NEXT: v_add_u16_e32 v32, 3, v32
|
|
; VI-NEXT: v_add_u16_e32 v17, 3, v17
|
|
; VI-NEXT: v_add_u16_e32 v31, 3, v31
|
|
; VI-NEXT: v_add_u16_e32 v18, 3, v18
|
|
; VI-NEXT: v_add_u16_e32 v30, 3, v30
|
|
; VI-NEXT: v_add_u16_e32 v19, 3, v19
|
|
; VI-NEXT: v_add_u16_e32 v29, 3, v29
|
|
; VI-NEXT: v_add_u16_e32 v20, 3, v20
|
|
; VI-NEXT: v_add_u16_e32 v28, 3, v28
|
|
; VI-NEXT: v_add_u16_e32 v21, 3, v21
|
|
; VI-NEXT: v_add_u16_e32 v27, 3, v27
|
|
; VI-NEXT: v_add_u16_e32 v22, 3, v22
|
|
; VI-NEXT: v_add_u16_e32 v26, 3, v26
|
|
; VI-NEXT: v_add_u16_e32 v23, 3, v23
|
|
; VI-NEXT: v_add_u16_e32 v25, 3, v25
|
|
; VI-NEXT: .LBB56_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v51
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v50
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v49
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v48
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v39
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v38
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v37
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v36
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v35
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v34
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v33
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v32
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v31
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v30
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v29
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v27
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v26
|
|
; 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_or_b32_sdwa v22, v22, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v25
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48i16_to_v48f16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v5
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v0
|
|
; 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 v23, v55, v23, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v54, v22, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v53, v21, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v52, v20, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v51, v19, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v50, v18, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v49, v17, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v48, v16, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v39, v15, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v38, v14, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v37, v13, s6
|
|
; GFX9-NEXT: v_perm_b32 v12, v36, v12, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v35, v11, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v34, v10, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v33, v9, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v32, v8, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v31, v7, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v30, v6, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v28, v5, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v29, v4, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v27, v3, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v26, v2, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v25, v1, s6
|
|
; GFX9-NEXT: v_perm_b32 v0, v24, v0, s6
|
|
; 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 v24, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v23
|
|
; 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 v0, v24, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v25, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v26, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v27, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v29, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v28, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v30, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v31, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v32, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v33, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v34, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v35, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v36, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v37, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v38, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v39, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v48, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v49, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v50, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v51, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v52, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v53, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v54, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v55, v23, s4
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v48f16:
|
|
; 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, v24
|
|
; 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 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_v48i16_to_v48f16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v18
|
|
; 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, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v0
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-FAKE16-NEXT: v_cmpx_ne_u32_e32 0, v24
|
|
; 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 v23, v64, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v55, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v54, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v53, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v52, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v51, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v50, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v49, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v39, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v38, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v37, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v35, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v34, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v33, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v31, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v30, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v25, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v26, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v27, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v28, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v29, v4, 0x5040100
|
|
; 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 v25, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v12
|
|
; 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, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v23
|
|
; GFX11-FAKE16-NEXT: .LBB56_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v25, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v26, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v27, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v28, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v29, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v30, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v31, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v33, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v34, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v35, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v37, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v38, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v39, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v49, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v50, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v51, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v52, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v53, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v54, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v55, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v64, v23, 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 <48 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <48 x i16> %a1 to <48 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x i16> %a to <48 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x half> %phi
|
|
}
|
|
|
|
define inreg <48 x half> @bitcast_v48i16_to_v48f16_scalar(<48 x i16> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v48i16_to_v48f16_scalar:
|
|
; 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:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:76 ; 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 v43, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: s_cbranch_scc0 .LBB57_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s22
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v41
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v60, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, s17
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s24
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v55
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, s18
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, s19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, s20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, s21
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s28
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, s23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, s25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, s27
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v14
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v17
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v20
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v23
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v26
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v30
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v40
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v54
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v7
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v8
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v9
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v10
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v11
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v12
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v28
|
|
; SI-NEXT: s_cbranch_execnz .LBB57_3
|
|
; SI-NEXT: .LBB57_2: ; %cmp.true
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 3, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
|
|
; SI-NEXT: v_add_i32_e32 v2, vcc, 3, v2
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 3, v3
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 3, v4
|
|
; SI-NEXT: v_add_i32_e32 v5, vcc, 3, v5
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 3, v6
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 3, v7
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v4
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 3, v8
|
|
; SI-NEXT: v_add_i32_e32 v9, vcc, 3, v9
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v5
|
|
; SI-NEXT: s_add_i32 s22, s22, 3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s22
|
|
; SI-NEXT: v_add_i32_e32 v10, vcc, 3, v10
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v6
|
|
; SI-NEXT: s_add_i32 s24, s24, 3
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s24
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v7
|
|
; SI-NEXT: v_add_i32_e32 v11, vcc, 3, v11
|
|
; SI-NEXT: s_add_i32 s26, s26, 3
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v8
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s26
|
|
; SI-NEXT: v_add_i32_e32 v12, vcc, 3, v12
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v9
|
|
; SI-NEXT: s_add_i32 s28, s28, 3
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s28
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v10
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_add_i32_e32 v41, vcc, 3, v41
|
|
; SI-NEXT: s_add_i32 s29, s29, 3
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v11
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, s29
|
|
; SI-NEXT: v_add_i32_e32 v54, vcc, 3, v54
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v12
|
|
; SI-NEXT: v_add_i32_e32 v55, vcc, 3, v55
|
|
; SI-NEXT: v_add_i32_e32 v40, vcc, 3, v40
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v41
|
|
; SI-NEXT: v_add_i32_e32 v30, vcc, 3, v30
|
|
; SI-NEXT: v_add_i32_e32 v29, vcc, 3, v29
|
|
; SI-NEXT: v_add_i32_e32 v28, vcc, 3, v28
|
|
; 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: s_add_i32 s27, s27, 3
|
|
; SI-NEXT: s_add_i32 s25, s25, 3
|
|
; SI-NEXT: s_add_i32 s23, s23, 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: v_cvt_f32_f16_e32 v60, s16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v35, s17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v61, s18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v36, s19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v45, s20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v62, s21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v57, s23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v46, s25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v42, s27
|
|
; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v50, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v53, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v48, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v52, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v38, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v51, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v34, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v49, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v21
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v39, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v63, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v37, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v58, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v47, v27
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v28
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v43, v29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v59, v30
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v56, v40
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v55
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v44, v54
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: .LBB57_3: ; %end
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v35
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v60
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v36
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v61
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v3
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
|
|
; SI-NEXT: v_or_b32_e32 v2, v4, v2
|
|
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v62
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v45
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v57
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v46
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v42
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v53
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v50
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v52
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v48
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v51
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v38
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v49
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v34
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v39
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v32
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v37
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v63
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v58
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v47
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v59
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v43
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v56
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB57_4:
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr60
|
|
; SI-NEXT: ; implicit-def: $vgpr35
|
|
; SI-NEXT: ; implicit-def: $vgpr61
|
|
; SI-NEXT: ; implicit-def: $vgpr36
|
|
; SI-NEXT: ; implicit-def: $vgpr45
|
|
; SI-NEXT: ; implicit-def: $vgpr62
|
|
; SI-NEXT: ; implicit-def: $vgpr57
|
|
; SI-NEXT: ; implicit-def: $vgpr46
|
|
; SI-NEXT: ; implicit-def: $vgpr42
|
|
; SI-NEXT: ; implicit-def: $vgpr50
|
|
; SI-NEXT: ; implicit-def: $vgpr53
|
|
; SI-NEXT: ; implicit-def: $vgpr48
|
|
; SI-NEXT: ; implicit-def: $vgpr52
|
|
; SI-NEXT: ; implicit-def: $vgpr38
|
|
; SI-NEXT: ; implicit-def: $vgpr51
|
|
; SI-NEXT: ; implicit-def: $vgpr34
|
|
; SI-NEXT: ; implicit-def: $vgpr49
|
|
; SI-NEXT: ; implicit-def: $vgpr32
|
|
; SI-NEXT: ; implicit-def: $vgpr39
|
|
; SI-NEXT: ; implicit-def: $vgpr63
|
|
; SI-NEXT: ; implicit-def: $vgpr37
|
|
; SI-NEXT: ; implicit-def: $vgpr58
|
|
; SI-NEXT: ; implicit-def: $vgpr33
|
|
; SI-NEXT: ; implicit-def: $vgpr47
|
|
; SI-NEXT: ; implicit-def: $vgpr43
|
|
; SI-NEXT: ; implicit-def: $vgpr59
|
|
; SI-NEXT: ; implicit-def: $vgpr56
|
|
; SI-NEXT: ; kill: killed $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr44
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: ; kill: killed $vgpr31
|
|
; SI-NEXT: ; implicit-def: $vgpr31
|
|
; SI-NEXT: s_branch .LBB57_2
|
|
;
|
|
; VI-LABEL: bitcast_v48i16_to_v48f16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v11, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v12, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v13, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v19, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v1
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v0
|
|
; 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 s43, s43, 3
|
|
; VI-NEXT: s_add_i32 s17, s17, 3
|
|
; VI-NEXT: s_add_i32 s42, s42, 3
|
|
; VI-NEXT: s_add_i32 s18, s18, 3
|
|
; VI-NEXT: s_add_i32 s41, s41, 3
|
|
; VI-NEXT: s_add_i32 s19, s19, 3
|
|
; VI-NEXT: s_add_i32 s40, s40, 3
|
|
; VI-NEXT: s_add_i32 s20, s20, 3
|
|
; VI-NEXT: s_add_i32 s15, s15, 3
|
|
; VI-NEXT: s_add_i32 s21, s21, 3
|
|
; VI-NEXT: s_add_i32 s14, s14, 3
|
|
; VI-NEXT: s_add_i32 s22, s22, 3
|
|
; VI-NEXT: s_add_i32 s13, s13, 3
|
|
; VI-NEXT: s_add_i32 s23, s23, 3
|
|
; VI-NEXT: s_add_i32 s12, s12, 3
|
|
; VI-NEXT: s_add_i32 s24, s24, 3
|
|
; VI-NEXT: s_add_i32 s11, s11, 3
|
|
; VI-NEXT: s_add_i32 s25, s25, 3
|
|
; VI-NEXT: s_add_i32 s10, s10, 3
|
|
; VI-NEXT: s_add_i32 s26, s26, 3
|
|
; VI-NEXT: s_add_i32 s9, s9, 3
|
|
; VI-NEXT: s_add_i32 s27, s27, 3
|
|
; VI-NEXT: s_add_i32 s8, s8, 3
|
|
; VI-NEXT: s_add_i32 s28, s28, 3
|
|
; VI-NEXT: s_add_i32 s7, s7, 3
|
|
; VI-NEXT: s_add_i32 s29, s29, 3
|
|
; VI-NEXT: s_add_i32 s6, s6, 3
|
|
; VI-NEXT: v_add_u32_e32 v0, vcc, 3, v0
|
|
; VI-NEXT: v_add_u32_e32 v10, vcc, 3, v10
|
|
; VI-NEXT: v_add_u32_e32 v1, vcc, 3, v1
|
|
; VI-NEXT: v_add_u32_e32 v15, vcc, 3, v15
|
|
; VI-NEXT: v_add_u32_e32 v2, vcc, 3, v2
|
|
; VI-NEXT: v_add_u32_e32 v16, vcc, 3, v16
|
|
; VI-NEXT: v_add_u32_e32 v3, vcc, 3, v3
|
|
; VI-NEXT: v_add_u32_e32 v17, vcc, 3, v17
|
|
; VI-NEXT: v_add_u32_e32 v4, vcc, 3, v4
|
|
; VI-NEXT: v_add_u32_e32 v18, vcc, 3, v18
|
|
; VI-NEXT: v_add_u32_e32 v5, vcc, 3, v5
|
|
; VI-NEXT: v_add_u32_e32 v19, vcc, 3, v19
|
|
; VI-NEXT: v_add_u32_e32 v6, vcc, 3, v6
|
|
; VI-NEXT: v_add_u32_e32 v20, vcc, 3, v20
|
|
; VI-NEXT: v_add_u32_e32 v7, vcc, 3, v7
|
|
; VI-NEXT: v_add_u32_e32 v13, vcc, 3, v13
|
|
; VI-NEXT: v_add_u32_e32 v8, vcc, 3, v8
|
|
; VI-NEXT: v_add_u32_e32 v12, vcc, 3, v12
|
|
; VI-NEXT: v_add_u32_e32 v9, vcc, 3, v9
|
|
; VI-NEXT: v_add_u32_e32 v11, vcc, 3, v11
|
|
; VI-NEXT: .LBB57_3: ; %end
|
|
; VI-NEXT: s_and_b32 s4, 0xffff, s16
|
|
; VI-NEXT: s_lshl_b32 s5, s43, 16
|
|
; VI-NEXT: s_or_b32 s4, s4, s5
|
|
; VI-NEXT: s_and_b32 s5, 0xffff, s17
|
|
; VI-NEXT: s_lshl_b32 s16, s42, 16
|
|
; VI-NEXT: s_or_b32 s5, s5, s16
|
|
; VI-NEXT: s_and_b32 s16, 0xffff, s18
|
|
; VI-NEXT: s_lshl_b32 s17, s41, 16
|
|
; VI-NEXT: s_or_b32 s16, s16, s17
|
|
; VI-NEXT: s_and_b32 s17, 0xffff, s19
|
|
; VI-NEXT: s_lshl_b32 s18, s40, 16
|
|
; VI-NEXT: s_or_b32 s17, s17, s18
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s20
|
|
; VI-NEXT: s_lshl_b32 s15, s15, 16
|
|
; VI-NEXT: s_or_b32 s15, s18, s15
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s21
|
|
; VI-NEXT: s_lshl_b32 s14, s14, 16
|
|
; VI-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; VI-NEXT: s_or_b32 s14, s18, s14
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s22
|
|
; VI-NEXT: s_lshl_b32 s13, s13, 16
|
|
; VI-NEXT: v_or_b32_sdwa v14, v0, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v15
|
|
; VI-NEXT: s_or_b32 s13, s18, s13
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s23
|
|
; VI-NEXT: s_lshl_b32 s12, s12, 16
|
|
; VI-NEXT: v_or_b32_sdwa v15, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v16
|
|
; VI-NEXT: s_or_b32 s12, s18, s12
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s24
|
|
; VI-NEXT: s_lshl_b32 s11, s11, 16
|
|
; VI-NEXT: v_or_b32_sdwa v16, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v17
|
|
; VI-NEXT: s_or_b32 s11, s18, s11
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s25
|
|
; VI-NEXT: s_lshl_b32 s10, s10, 16
|
|
; VI-NEXT: v_or_b32_sdwa v17, v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v18
|
|
; VI-NEXT: s_or_b32 s10, s18, s10
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s26
|
|
; VI-NEXT: s_lshl_b32 s9, s9, 16
|
|
; VI-NEXT: v_or_b32_sdwa v18, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v19
|
|
; VI-NEXT: s_or_b32 s9, s18, s9
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s27
|
|
; VI-NEXT: s_lshl_b32 s8, s8, 16
|
|
; VI-NEXT: v_or_b32_sdwa v19, v5, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v20
|
|
; VI-NEXT: s_or_b32 s8, s18, s8
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s28
|
|
; VI-NEXT: s_lshl_b32 s7, s7, 16
|
|
; VI-NEXT: v_or_b32_sdwa v20, v6, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v13
|
|
; VI-NEXT: s_or_b32 s7, s18, s7
|
|
; VI-NEXT: s_and_b32 s18, 0xffff, s29
|
|
; VI-NEXT: s_lshl_b32 s6, s6, 16
|
|
; VI-NEXT: v_or_b32_sdwa v21, v7, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v12
|
|
; VI-NEXT: s_or_b32 s6, s18, s6
|
|
; VI-NEXT: v_or_b32_sdwa v22, v8, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v11
|
|
; VI-NEXT: v_or_b32_sdwa v23, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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, s15
|
|
; VI-NEXT: v_mov_b32_e32 v5, s14
|
|
; VI-NEXT: v_mov_b32_e32 v6, s13
|
|
; VI-NEXT: v_mov_b32_e32 v7, s12
|
|
; VI-NEXT: v_mov_b32_e32 v8, s11
|
|
; VI-NEXT: v_mov_b32_e32 v9, s10
|
|
; VI-NEXT: v_mov_b32_e32 v10, s9
|
|
; VI-NEXT: v_mov_b32_e32 v11, s8
|
|
; VI-NEXT: v_mov_b32_e32 v12, s7
|
|
; VI-NEXT: v_mov_b32_e32 v13, s6
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
; VI-NEXT: .LBB57_4:
|
|
; VI-NEXT: s_branch .LBB57_2
|
|
;
|
|
; GFX9-LABEL: bitcast_v48i16_to_v48f16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v1
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v0
|
|
; 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, 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 v29, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s24, s14
|
|
; GFX9-NEXT: v_pk_add_u16 v28, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s23, s13
|
|
; GFX9-NEXT: v_pk_add_u16 v27, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s22, s12
|
|
; GFX9-NEXT: v_pk_add_u16 v26, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s21, s11
|
|
; GFX9-NEXT: v_pk_add_u16 v25, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s20, s10
|
|
; GFX9-NEXT: v_pk_add_u16 v24, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s19, s9
|
|
; GFX9-NEXT: v_pk_add_u16 v33, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s18, s8
|
|
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_pk_add_u16 v32, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s17, s7
|
|
; GFX9-NEXT: v_lshl_or_b32 v9, v23, 16, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v8, v22, 16, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v7, v21, 16, v7
|
|
; GFX9-NEXT: v_lshl_or_b32 v6, v20, 16, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v5, v19, 16, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v4, v18, 16, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v3, v17, 16, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v2, v16, 16, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v1, v15, 16, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v0, v14, 16, v0
|
|
; GFX9-NEXT: v_pk_add_u16 v31, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s6
|
|
; 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_pk_add_u16 v30, s4, 3 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v30
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v31
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v32
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v33
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v28
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v29
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v9
|
|
; 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 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 v29, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, s6
|
|
; GFX9-NEXT: .LBB57_5: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v14, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v15, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v16, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v17, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v18, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v19, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v20, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v7
|
|
; GFX9-NEXT: v_and_b32_e32 v30, 0xffff, v30
|
|
; GFX9-NEXT: v_and_b32_e32 v31, 0xffff, v31
|
|
; GFX9-NEXT: v_and_b32_e32 v32, 0xffff, v32
|
|
; GFX9-NEXT: v_and_b32_e32 v33, 0xffff, v33
|
|
; 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_and_b32_e32 v28, 0xffff, v28
|
|
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff, v29
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v21, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v30, v55, 16, v30
|
|
; GFX9-NEXT: v_lshl_or_b32 v31, v54, 16, v31
|
|
; GFX9-NEXT: v_lshl_or_b32 v32, v53, 16, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v33, v52, 16, v33
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v51, 16, v24
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v50, 16, v25
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v49, 16, v26
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v48, 16, v27
|
|
; GFX9-NEXT: v_lshl_or_b32 v28, v39, 16, v28
|
|
; GFX9-NEXT: v_lshl_or_b32 v29, v38, 16, v29
|
|
; 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_lshl_or_b32 v22, v22, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v37, 16, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v36, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v35, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v34, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v23, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, v30
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, v31
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, v32
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, v25
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, v26
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, v27
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, v28
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, v29
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v48f16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0
|
|
; 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 s13, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, 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 s14, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s46, 0
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB57_3
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v23.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v24.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v21.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v24.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v20.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v24.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v19.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v24.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v18.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v24.h
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s46
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB57_4
|
|
; GFX11-TRUE16-NEXT: .LBB57_2: ; %cmp.true
|
|
; 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 s13, s22, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s21, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s20, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s19, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s18, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s17, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s4
|
|
; 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, 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, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s8, 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 v55, 16, v0
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v23
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB57_5
|
|
; GFX11-TRUE16-NEXT: .LBB57_3:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB57_2
|
|
; GFX11-TRUE16-NEXT: .LBB57_4:
|
|
; 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 v30, s45 :: v_dual_mov_b32 v31, s44
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v32, s43 :: v_dual_mov_b32 v33, s42
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v34, s41 :: v_dual_mov_b32 v35, s40
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v36, s15 :: v_dual_mov_b32 v37, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v38, s12 :: v_dual_mov_b32 v39, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v48, s10 :: v_dual_mov_b32 v49, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v50, s8 :: v_dual_mov_b32 v51, s6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v52, s5 :: v_dual_mov_b32 v53, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s4 :: v_dual_mov_b32 v55, s14
|
|
; GFX11-TRUE16-NEXT: .LBB57_5: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v55, v55 :: v_dual_mov_b32 v54, v54
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v53 :: v_dual_mov_b32 v52, v52
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v51, v51 :: v_dual_mov_b32 v50, v50
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v49, v49 :: v_dual_mov_b32 v48, v48
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, v39 :: v_dual_mov_b32 v38, v38
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, v37 :: v_dual_mov_b32 v36, v36
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, v35 :: v_dual_mov_b32 v34, v34
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, v33 :: v_dual_mov_b32 v32, v32
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, v31 :: v_dual_mov_b32 v30, v30
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v29 :: v_dual_mov_b32 v28, v28
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v27 :: v_dual_mov_b32 v26, v26
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v25 :: v_dual_mov_b32 v24, v24
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v31.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v28.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v27.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v26.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v25.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v24.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v48i16_to_v48f16_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v19, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v18, 16, v0
|
|
; 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 s7, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s46, 0
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; 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, s46
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB57_4
|
|
; GFX11-FAKE16-NEXT: .LBB57_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; 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 v0, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v23, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v22, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v21, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v20, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v19, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v18, 16, v0
|
|
; 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 s7, s16, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s8
|
|
; 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 v5, v5, 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 v3, v3, 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 v1, v1, 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 v14, s29, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v15, s28, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v16, s27, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v17, s26, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v9, s25, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v10, s24, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v11, s15, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v12, s14, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v6, s12, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v7, s11, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v8, s10, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v28, s9, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v27, s0, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v26, s1, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v25, s2, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v24, s3, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_u16 v29, s7, 3 op_sel_hi:[1,0]
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v28
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v18, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v19, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v5
|
|
; 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 v14, s29 :: v_dual_mov_b32 v15, s28
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s27 :: v_dual_mov_b32 v17, s26
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s25 :: v_dual_mov_b32 v10, s24
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v12, s22
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s21 :: v_dual_mov_b32 v6, s20
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v8, s18
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v28, s17 :: v_dual_mov_b32 v29, s16
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s3 :: v_dual_mov_b32 v25, s2
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v26, s1 :: v_dual_mov_b32 v27, s0
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v30, s45 :: v_dual_mov_b32 v31, s44
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v32, s43 :: v_dual_mov_b32 v33, s42
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v34, s41 :: v_dual_mov_b32 v35, s40
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, s15 :: v_dual_mov_b32 v37, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v38, s13 :: v_dual_mov_b32 v39, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v48, s11 :: v_dual_mov_b32 v49, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v50, s9 :: v_dual_mov_b32 v51, s7
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v52, s6 :: v_dual_mov_b32 v53, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v54, s4 :: v_dual_mov_b32 v55, s5
|
|
; GFX11-FAKE16-NEXT: .LBB57_5: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v65, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v55, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v52, 16, v65
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xffff, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v18, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v54, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v28
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v22, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, v27
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v53, 16, v64
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v53, 0xffff, v6
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v49, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v39, 16, v53
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v28, v51, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v29, v50, 16, v52
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v38, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v37, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v35, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v15
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v33, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v31, 16, v34
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v30, 16, v35
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v19, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v20, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v21, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v23, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, v26
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, v28 :: v_dual_mov_b32 v5, 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 = add <48 x i16> %a, splat (i16 3)
|
|
%a2 = bitcast <48 x i16> %a1 to <48 x half>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x i16> %a to <48 x half>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x half> %phi
|
|
}
|
|
|
|
define <48 x i16> @bitcast_v48f16_to_v48i16(<48 x half> %a, i32 %b) {
|
|
; SI-LABEL: bitcast_v48f16_to_v48i16:
|
|
; 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:136 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: s_waitcnt expcnt(6)
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: s_waitcnt expcnt(5)
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:20
|
|
; SI-NEXT: s_waitcnt expcnt(4)
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:24
|
|
; SI-NEXT: s_waitcnt expcnt(3)
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:28
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:32
|
|
; SI-NEXT: s_waitcnt expcnt(1)
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:36
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:40
|
|
; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:44
|
|
; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:48
|
|
; SI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:52
|
|
; SI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:56
|
|
; SI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:60
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:64
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
|
|
; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:72
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v37, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v40, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v46
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, v47
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v56
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v57
|
|
; SI-NEXT: s_waitcnt vmcnt(13)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v58
|
|
; SI-NEXT: s_waitcnt vmcnt(12)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v57, v59
|
|
; SI-NEXT: s_waitcnt vmcnt(11)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v60
|
|
; SI-NEXT: s_waitcnt vmcnt(10)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v61
|
|
; SI-NEXT: s_waitcnt vmcnt(9)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v61, v62
|
|
; SI-NEXT: s_waitcnt vmcnt(8)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v56, v63
|
|
; SI-NEXT: s_waitcnt vmcnt(7)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v23, v34
|
|
; SI-NEXT: s_waitcnt vmcnt(6)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v35
|
|
; SI-NEXT: s_waitcnt vmcnt(5)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v49
|
|
; SI-NEXT: s_waitcnt vmcnt(4)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, v52
|
|
; SI-NEXT: s_waitcnt vmcnt(3)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v53
|
|
; SI-NEXT: s_waitcnt vmcnt(2)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v59, v31
|
|
; SI-NEXT: s_waitcnt vmcnt(1)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v32
|
|
; SI-NEXT: s_waitcnt vmcnt(0)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v33
|
|
; 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_or_saveexec_b64 s[4:5], s[4:5]
|
|
; SI-NEXT: v_mov_b32_e32 v52, v21
|
|
; SI-NEXT: v_mov_b32_e32 v21, v6
|
|
; SI-NEXT: v_mov_b32_e32 v6, v7
|
|
; SI-NEXT: v_mov_b32_e32 v7, v8
|
|
; SI-NEXT: v_mov_b32_e32 v8, v1
|
|
; SI-NEXT: v_mov_b32_e32 v58, v2
|
|
; SI-NEXT: v_mov_b32_e32 v60, v50
|
|
; SI-NEXT: v_mov_b32_e32 v50, v3
|
|
; SI-NEXT: s_xor_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: s_cbranch_execz .LBB58_2
|
|
; SI-NEXT: ; %bb.1: ; %cmp.true
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v27, v44
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v60
|
|
; SI-NEXT: v_mov_b32_e32 v35, v29
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v29, v43
|
|
; 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 v27, 16, v27
|
|
; SI-NEXT: v_or_b32_e32 v60, v28, v27
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v28, v42
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v55
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_add_f32_e32 v28, 0x38000000, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v28
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v29
|
|
; SI-NEXT: v_add_f32_e32 v32, 0x38000000, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v40
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v32
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v52
|
|
; SI-NEXT: v_add_f32_e32 v31, 0x38000000, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_add_f32_e32 v32, 0x38000000, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v55, 16, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v54
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v31, 0x38000000, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v53
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v59
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v31
|
|
; SI-NEXT: v_or_b32_e32 v52, v32, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v35
|
|
; SI-NEXT: v_add_f32_e32 v32, 0x38000000, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v33, v25
|
|
; SI-NEXT: v_add_f32_e32 v31, 0x38000000, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v32
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v46
|
|
; SI-NEXT: v_lshlrev_b32_e32 v46, 16, v33
|
|
; SI-NEXT: v_or_b32_e32 v49, v31, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v31, 0x38000000, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v57
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v61
|
|
; SI-NEXT: v_or_b32_e32 v35, v31, v46
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v23
|
|
; SI-NEXT: v_add_f32_e32 v32, 0x38000000, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v56
|
|
; SI-NEXT: v_add_f32_e32 v31, 0x38000000, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_or_b32_e32 v57, v32, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v22, 16, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v34
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
|
|
; SI-NEXT: v_or_b32_e32 v56, v24, v22
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v31
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v1
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v61, v3
|
|
; SI-NEXT: v_add_f32_e32 v11, 0x38000000, v11
|
|
; SI-NEXT: v_add_f32_e32 v9, 0x38000000, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v47, 16, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_add_f32_e32 v13, 0x38000000, v13
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
|
|
; SI-NEXT: v_or_b32_e32 v34, v24, v47
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v53
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v38
|
|
; SI-NEXT: v_or_b32_e32 v59, v2, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_or_b32_e32 v5, v5, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v61
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_add_f32_e32 v19, 0x38000000, v19
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_or_b32_e32 v11, v11, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v17, 0x38000000, v17
|
|
; SI-NEXT: v_add_f32_e32 v15, 0x38000000, v15
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v36
|
|
; SI-NEXT: v_add_f32_e32 v31, 0x38000000, v31
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v12
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v33, v37
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v58
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v19, v19, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v17, v17, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v24, 16, v15
|
|
; SI-NEXT: v_add_f32_e32 v32, 0x38000000, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v32
|
|
; SI-NEXT: v_add_f32_e32 v32, 0x38000000, v33
|
|
; SI-NEXT: v_or_b32_e32 v38, v31, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v39
|
|
; SI-NEXT: v_add_f32_e32 v30, 0x38000000, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v30, v30
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v36
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_or_b32_e32 v37, v32, v31
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v31, v51
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v24
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v48
|
|
; SI-NEXT: v_or_b32_e32 v58, v30, v28
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v30, v41
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v32, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
|
|
; SI-NEXT: v_add_f32_e32 v31, 0x38000000, v31
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_add_f32_e32 v30, 0x38000000, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v24
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v32
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v30, v30
|
|
; SI-NEXT: v_add_f32_e32 v7, 0x38000000, v7
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_add_f32_e32 v21, 0x38000000, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v51, v31, v32
|
|
; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v48
|
|
; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
|
|
; SI-NEXT: v_or_b32_e32 v50, v24, v31
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v29
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v30
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v55
|
|
; SI-NEXT: v_or_b32_e32 v21, v21, v45
|
|
; SI-NEXT: v_alignbit_b32 v44, v50, v27, 16
|
|
; SI-NEXT: v_alignbit_b32 v43, v51, v28, 16
|
|
; SI-NEXT: v_alignbit_b32 v42, v37, v29, 16
|
|
; SI-NEXT: v_mov_b32_e32 v29, v49
|
|
; SI-NEXT: v_alignbit_b32 v41, v38, v30, 16
|
|
; SI-NEXT: v_alignbit_b32 v40, v17, v55, 16
|
|
; SI-NEXT: v_alignbit_b32 v55, v19, v45, 16
|
|
; SI-NEXT: v_alignbit_b32 v54, v20, v26, 16
|
|
; SI-NEXT: v_alignbit_b32 v26, v14, v25, 16
|
|
; SI-NEXT: v_alignbit_b32 v25, v10, v46, 16
|
|
; SI-NEXT: v_mov_b32_e32 v46, v35
|
|
; SI-NEXT: v_alignbit_b32 v24, v11, v23, 16
|
|
; SI-NEXT: v_alignbit_b32 v23, v5, v22, 16
|
|
; SI-NEXT: v_alignbit_b32 v22, v59, v47, 16
|
|
; SI-NEXT: .LBB58_2: ; %end
|
|
; SI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v60
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v44
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v50
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: v_add_i32_e32 v28, vcc, 4, v0
|
|
; SI-NEXT: buffer_store_dword v27, v28, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v58
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: v_add_i32_e32 v28, vcc, 8, v0
|
|
; SI-NEXT: buffer_store_dword v27, v28, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v27, 0xffff, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v28, 16, v39
|
|
; SI-NEXT: v_or_b32_e32 v27, v27, v28
|
|
; SI-NEXT: v_add_i32_e32 v28, vcc, 12, v0
|
|
; SI-NEXT: buffer_store_dword v27, v28, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v27
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 16, v0
|
|
; SI-NEXT: buffer_store_dword v8, v27, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v8, 0xffff, v37
|
|
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v27
|
|
; SI-NEXT: v_add_i32_e32 v27, vcc, 20, v0
|
|
; SI-NEXT: buffer_store_dword v8, v27, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v41
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 24, v0
|
|
; SI-NEXT: buffer_store_dword v7, v8, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v7, 0xffff, v38
|
|
; SI-NEXT: v_lshlrev_b32_e32 v8, 16, v15
|
|
; SI-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; SI-NEXT: v_add_i32_e32 v8, vcc, 28, v0
|
|
; SI-NEXT: buffer_store_dword v7, v8, s[0:3], 0 offen
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 32, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v17
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v16
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 36, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v21
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v55
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 40, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v19
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v18
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 44, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v52
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v54
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 48, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v12
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 52, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v29
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v26
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 56, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v13
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 60, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v46
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 64, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v9
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 0x44, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v57
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v24
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; SI-NEXT: v_add_i32_e32 v7, vcc, 0x48, v0
|
|
; SI-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v6, 0xffff, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v61
|
|
; SI-NEXT: v_or_b32_e32 v3, v6, v3
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 0x4c, v0
|
|
; SI-NEXT: buffer_store_dword v3, v6, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v23
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v6
|
|
; SI-NEXT: v_add_i32_e32 v6, vcc, 0x50, v0
|
|
; SI-NEXT: buffer_store_dword v3, v6, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 0x54, v0
|
|
; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v34
|
|
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v22
|
|
; SI-NEXT: v_and_b32_e32 v2, 0xffff, v59
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v53
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; SI-NEXT: v_add_i32_e32 v4, vcc, 0x58, v0
|
|
; SI-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-LABEL: bitcast_v48f16_to_v48i16:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-NEXT: v_lshrrev_b32_e32 v25, 16, v23
|
|
; VI-NEXT: v_lshrrev_b32_e32 v26, 16, v22
|
|
; VI-NEXT: v_lshrrev_b32_e32 v27, 16, v21
|
|
; VI-NEXT: v_lshrrev_b32_e32 v28, 16, v20
|
|
; VI-NEXT: v_lshrrev_b32_e32 v29, 16, v19
|
|
; VI-NEXT: v_lshrrev_b32_e32 v30, 16, v18
|
|
; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v17
|
|
; VI-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; VI-NEXT: v_lshrrev_b32_e32 v33, 16, v15
|
|
; VI-NEXT: v_lshrrev_b32_e32 v34, 16, v14
|
|
; VI-NEXT: v_lshrrev_b32_e32 v35, 16, v13
|
|
; VI-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; VI-NEXT: v_lshrrev_b32_e32 v37, 16, v11
|
|
; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v39, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v48, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v49, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v50, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v51, 16, v5
|
|
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; VI-NEXT: v_lshrrev_b32_e32 v24, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v0
|
|
; 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 v55, 0x200, v55
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
|
|
; VI-NEXT: v_add_f16_e32 v54, 0x200, v54
|
|
; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
|
|
; VI-NEXT: v_add_f16_e32 v53, 0x200, v53
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
|
|
; VI-NEXT: v_add_f16_e32 v52, 0x200, v52
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
|
|
; VI-NEXT: v_add_f16_e32 v24, 0x200, v24
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
|
|
; VI-NEXT: v_add_f16_e32 v51, 0x200, v51
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
|
|
; VI-NEXT: v_add_f16_e32 v50, 0x200, v50
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
|
|
; VI-NEXT: v_add_f16_e32 v49, 0x200, v49
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
|
|
; VI-NEXT: v_add_f16_e32 v48, 0x200, v48
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
|
|
; VI-NEXT: v_add_f16_e32 v39, 0x200, v39
|
|
; VI-NEXT: v_add_f16_e32 v10, 0x200, v10
|
|
; VI-NEXT: v_add_f16_e32 v38, 0x200, v38
|
|
; VI-NEXT: v_add_f16_e32 v11, 0x200, v11
|
|
; VI-NEXT: v_add_f16_e32 v37, 0x200, v37
|
|
; VI-NEXT: v_add_f16_e32 v12, 0x200, v12
|
|
; VI-NEXT: v_add_f16_e32 v36, 0x200, v36
|
|
; VI-NEXT: v_add_f16_e32 v13, 0x200, v13
|
|
; VI-NEXT: v_add_f16_e32 v35, 0x200, v35
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
|
|
; VI-NEXT: v_add_f16_e32 v34, 0x200, v34
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
|
|
; VI-NEXT: v_add_f16_e32 v33, 0x200, v33
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
|
|
; VI-NEXT: v_add_f16_e32 v32, 0x200, v32
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v17
|
|
; VI-NEXT: v_add_f16_e32 v31, 0x200, v31
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v18
|
|
; VI-NEXT: v_add_f16_e32 v30, 0x200, v30
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v19
|
|
; VI-NEXT: v_add_f16_e32 v29, 0x200, v29
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v20
|
|
; VI-NEXT: v_add_f16_e32 v28, 0x200, v28
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v21
|
|
; VI-NEXT: v_add_f16_e32 v27, 0x200, v27
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v22
|
|
; VI-NEXT: v_add_f16_e32 v26, 0x200, v26
|
|
; VI-NEXT: v_add_f16_e32 v23, 0x200, v23
|
|
; VI-NEXT: v_add_f16_e32 v25, 0x200, v25
|
|
; VI-NEXT: .LBB58_2: ; %end
|
|
; VI-NEXT: s_or_b64 exec, exec, s[4:5]
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v24
|
|
; VI-NEXT: v_or_b32_sdwa v4, v4, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v51
|
|
; VI-NEXT: v_or_b32_sdwa v5, v5, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v50
|
|
; VI-NEXT: v_or_b32_sdwa v6, v6, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v49
|
|
; VI-NEXT: v_or_b32_sdwa v7, v7, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v48
|
|
; VI-NEXT: v_or_b32_sdwa v8, v8, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v39
|
|
; VI-NEXT: v_or_b32_sdwa v9, v9, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v38
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v37
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v36
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v35
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v34
|
|
; VI-NEXT: v_or_b32_sdwa v14, v14, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v33
|
|
; VI-NEXT: v_or_b32_sdwa v15, v15, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v32
|
|
; VI-NEXT: v_or_b32_sdwa v16, v16, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v31
|
|
; VI-NEXT: v_or_b32_sdwa v17, v17, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v30
|
|
; VI-NEXT: v_or_b32_sdwa v18, v18, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v29
|
|
; VI-NEXT: v_or_b32_sdwa v19, v19, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v28
|
|
; VI-NEXT: v_or_b32_sdwa v20, v20, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v27
|
|
; VI-NEXT: v_or_b32_sdwa v21, v21, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v26
|
|
; 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_or_b32_sdwa v22, v22, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v24, 16, v25
|
|
; VI-NEXT: v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v3, v3, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48f16_to_v48i16:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v23
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v5
|
|
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v0
|
|
; 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 v23, v55, v23, s6
|
|
; GFX9-NEXT: v_perm_b32 v22, v54, v22, s6
|
|
; GFX9-NEXT: v_perm_b32 v21, v53, v21, s6
|
|
; GFX9-NEXT: v_perm_b32 v20, v52, v20, s6
|
|
; GFX9-NEXT: v_perm_b32 v19, v51, v19, s6
|
|
; GFX9-NEXT: v_perm_b32 v18, v50, v18, s6
|
|
; GFX9-NEXT: v_perm_b32 v17, v49, v17, s6
|
|
; GFX9-NEXT: v_perm_b32 v16, v48, v16, s6
|
|
; GFX9-NEXT: v_perm_b32 v15, v39, v15, s6
|
|
; GFX9-NEXT: v_perm_b32 v14, v38, v14, s6
|
|
; GFX9-NEXT: v_perm_b32 v13, v37, v13, s6
|
|
; GFX9-NEXT: s_movk_i32 s7, 0x200
|
|
; GFX9-NEXT: v_perm_b32 v12, v36, v12, s6
|
|
; GFX9-NEXT: v_perm_b32 v11, v35, v11, s6
|
|
; GFX9-NEXT: v_perm_b32 v10, v34, v10, s6
|
|
; GFX9-NEXT: v_perm_b32 v9, v33, v9, s6
|
|
; GFX9-NEXT: v_perm_b32 v8, v32, v8, s6
|
|
; GFX9-NEXT: v_perm_b32 v7, v31, v7, s6
|
|
; GFX9-NEXT: v_perm_b32 v6, v30, v6, s6
|
|
; GFX9-NEXT: v_perm_b32 v5, v28, v5, s6
|
|
; GFX9-NEXT: v_perm_b32 v4, v29, v4, s6
|
|
; GFX9-NEXT: v_perm_b32 v3, v27, v3, s6
|
|
; GFX9-NEXT: v_perm_b32 v2, v26, v2, s6
|
|
; GFX9-NEXT: v_perm_b32 v1, v25, v1, s6
|
|
; GFX9-NEXT: v_perm_b32 v0, v24, v0, s6
|
|
; 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 v24, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v29, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v33, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v14
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v15
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v16
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v17
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v18
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v19
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v20
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v21
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v22
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v23
|
|
; 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 v0, v24, v0, s4
|
|
; GFX9-NEXT: v_perm_b32 v1, v25, v1, s4
|
|
; GFX9-NEXT: v_perm_b32 v2, v26, v2, s4
|
|
; GFX9-NEXT: v_perm_b32 v3, v27, v3, s4
|
|
; GFX9-NEXT: v_perm_b32 v4, v29, v4, s4
|
|
; GFX9-NEXT: v_perm_b32 v5, v28, v5, s4
|
|
; GFX9-NEXT: v_perm_b32 v6, v30, v6, s4
|
|
; GFX9-NEXT: v_perm_b32 v7, v31, v7, s4
|
|
; GFX9-NEXT: v_perm_b32 v8, v32, v8, s4
|
|
; GFX9-NEXT: v_perm_b32 v9, v33, v9, s4
|
|
; GFX9-NEXT: v_perm_b32 v10, v34, v10, s4
|
|
; GFX9-NEXT: v_perm_b32 v11, v35, v11, s4
|
|
; GFX9-NEXT: v_perm_b32 v12, v36, v12, s4
|
|
; GFX9-NEXT: v_perm_b32 v13, v37, v13, s4
|
|
; GFX9-NEXT: v_perm_b32 v14, v38, v14, s4
|
|
; GFX9-NEXT: v_perm_b32 v15, v39, v15, s4
|
|
; GFX9-NEXT: v_perm_b32 v16, v48, v16, s4
|
|
; GFX9-NEXT: v_perm_b32 v17, v49, v17, s4
|
|
; GFX9-NEXT: v_perm_b32 v18, v50, v18, s4
|
|
; GFX9-NEXT: v_perm_b32 v19, v51, v19, s4
|
|
; GFX9-NEXT: v_perm_b32 v20, v52, v20, s4
|
|
; GFX9-NEXT: v_perm_b32 v21, v53, v21, s4
|
|
; GFX9-NEXT: v_perm_b32 v22, v54, v22, s4
|
|
; GFX9-NEXT: v_perm_b32 v23, v55, v23, s4
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v48i16:
|
|
; 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, v24
|
|
; 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 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_v48f16_to_v48i16:
|
|
; GFX11-FAKE16: ; %bb.0:
|
|
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v23
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v18
|
|
; 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, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v25, 16, v0
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s0, exec_lo
|
|
; GFX11-FAKE16-NEXT: v_cmpx_ne_u32_e32 0, v24
|
|
; 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 v23, v64, v23, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v55, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v54, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v53, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v52, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v51, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v50, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v49, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v39, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v38, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v37, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v35, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v34, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v33, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v31, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v30, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v25, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v26, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v27, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v28, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v29, v4, 0x5040100
|
|
; 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 v25, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v26, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v27, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v28, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v29, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v12
|
|
; 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, v18
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v19
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v20
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v21
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v22
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v23
|
|
; GFX11-FAKE16-NEXT: .LBB58_2: ; %end
|
|
; GFX11-FAKE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v0, v25, v0, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v1, v26, v1, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v2, v27, v2, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v3, v28, v3, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v4, v29, v4, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v5, v30, v5, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v6, v31, v6, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v7, v32, v7, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v8, v33, v8, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v9, v34, v9, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v10, v35, v10, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v11, v36, v11, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v12, v37, v12, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v13, v38, v13, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v14, v39, v14, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v15, v48, v15, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v16, v49, v16, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v17, v50, v17, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v18, v51, v18, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v19, v52, v19, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v20, v53, v20, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v21, v54, v21, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v22, v55, v22, 0x5040100
|
|
; GFX11-FAKE16-NEXT: v_perm_b32 v23, v64, v23, 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 <48 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <48 x half> %a1 to <48 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x half> %a to <48 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x i16> %phi
|
|
}
|
|
|
|
define inreg <48 x i16> @bitcast_v48f16_to_v48i16_scalar(<48 x half> inreg %a, i32 inreg %b) {
|
|
; SI-LABEL: bitcast_v48f16_to_v48i16_scalar:
|
|
; SI: ; %bb.0:
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:16
|
|
; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32
|
|
; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:12
|
|
; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:4
|
|
; SI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:8
|
|
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:76 ; 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 v43, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
|
|
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(2)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v61, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v55, v3
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v63, v6
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v44, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v60, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v53, v15
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v56, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v51, v19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v45, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v49, v23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v40, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v39, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v30
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v62, s16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, s19
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v59, s20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v58, s23
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v57, s24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, s27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, s28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v48, v4
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v38, v8
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v9
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v36, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v13
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v34, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v30, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v27, v28
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v29
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v54, s17
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, s18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v52, s21
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v50, s25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, s26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v29, s29
|
|
; SI-NEXT: s_waitcnt vmcnt(14)
|
|
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v31, v32
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v33
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v32, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v35
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v37
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, s22
|
|
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; SI-NEXT: s_cbranch_scc0 .LBB59_4
|
|
; SI-NEXT: ; %bb.1: ; %cmp.false
|
|
; SI-NEXT: s_cbranch_execnz .LBB59_3
|
|
; SI-NEXT: .LBB59_2: ; %cmp.true
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v54
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v52
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v50
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v7, v29
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v23, 16, 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, v7
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v1
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v38
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v57
|
|
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v48
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
|
|
; SI-NEXT: v_add_f32_e32 v2, 0x38000000, v2
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v1
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v59
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
|
|
; SI-NEXT: v_add_f32_e32 v6, 0x38000000, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v62
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_or_b32_e32 v59, v3, v19
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v5
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_or_b32_e32 v62, v1, v23
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v55
|
|
; SI-NEXT: v_or_b32_e32 v57, v3, v15
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v44
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; 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_or_b32_e32 v1, v1, v21
|
|
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v36
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_or_b32_e32 v35, v5, v17
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v11
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_or_b32_e32 v3, v3, v13
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v34
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v53
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_or_b32_e32 v33, v5, v11
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v32
|
|
; 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 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v51
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
|
|
; SI-NEXT: v_or_b32_e32 v32, v1, v9
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v30
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v5
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v5, v49
|
|
; 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 v1, v1
|
|
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
|
|
; SI-NEXT: v_or_b32_e32 v29, v3, v7
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v3, v27
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v5
|
|
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v1, v39
|
|
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
|
|
; SI-NEXT: v_or_b32_e32 v28, v26, v5
|
|
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v31
|
|
; SI-NEXT: v_add_f32_e32 v8, 0x38000000, v8
|
|
; SI-NEXT: v_or_b32_e32 v27, v1, v3
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v43
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v43, v25
|
|
; SI-NEXT: v_or_b32_e32 v44, v26, v1
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v41
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v2, v2, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v40
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v41, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v45
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v40, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v41
|
|
; SI-NEXT: v_or_b32_e32 v4, v4, v25
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v45, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v6, v6, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v56
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v8, v8, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v47
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v56, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v60
|
|
; SI-NEXT: v_add_f32_e32 v10, 0x38000000, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v47, v25
|
|
; SI-NEXT: v_add_f32_e32 v12, 0x38000000, v12
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v60, v26
|
|
; SI-NEXT: v_add_f32_e32 v14, 0x38000000, v14
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v56
|
|
; SI-NEXT: v_or_b32_e32 v10, v10, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v12, v12, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v63
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v60
|
|
; SI-NEXT: v_or_b32_e32 v14, v14, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v61
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v63, v25
|
|
; SI-NEXT: v_add_f32_e32 v18, 0x38000000, v18
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v61, v25
|
|
; SI-NEXT: v_add_f32_e32 v22, 0x38000000, v22
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v63
|
|
; SI-NEXT: v_or_b32_e32 v18, v18, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v61
|
|
; SI-NEXT: v_or_b32_e32 v22, v22, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v58
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v26, v46
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v58, v25
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v25, v42
|
|
; SI-NEXT: v_add_f32_e32 v26, 0x38000000, v26
|
|
; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v46, v26
|
|
; SI-NEXT: v_add_f32_e32 v16, 0x38000000, v16
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
|
|
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
|
|
; SI-NEXT: v_add_f32_e32 v25, 0x38000000, v25
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v42, v25
|
|
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v26, 16, v46
|
|
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
|
|
; SI-NEXT: v_or_b32_e32 v16, v16, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v58
|
|
; SI-NEXT: v_lshr_b64 v[50:51], v[15:16], 16
|
|
; SI-NEXT: v_or_b32_e32 v20, v20, v25
|
|
; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v42
|
|
; SI-NEXT: v_mov_b32_e32 v51, v29
|
|
; SI-NEXT: v_lshr_b64 v[29:30], v[21:22], 16
|
|
; SI-NEXT: v_or_b32_e32 v24, v24, v25
|
|
; SI-NEXT: v_lshr_b64 v[30:31], v[5:6], 16
|
|
; SI-NEXT: v_lshr_b64 v[25:26], v[1:2], 16
|
|
; SI-NEXT: v_mov_b32_e32 v31, v44
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
|
|
; SI-NEXT: v_lshr_b64 v[54:55], v[23:24], 16
|
|
; SI-NEXT: v_lshr_b64 v[52:53], v[19:20], 16
|
|
; SI-NEXT: v_lshr_b64 v[48:49], v[17:18], 16
|
|
; SI-NEXT: v_lshr_b64 v[38:39], v[13:14], 16
|
|
; SI-NEXT: v_mov_b32_e32 v55, v35
|
|
; SI-NEXT: v_mov_b32_e32 v53, v32
|
|
; SI-NEXT: v_mov_b32_e32 v49, v28
|
|
; SI-NEXT: v_mov_b32_e32 v39, v27
|
|
; SI-NEXT: v_lshr_b64 v[36:37], v[11:12], 16
|
|
; SI-NEXT: v_mov_b32_e32 v11, v33
|
|
; SI-NEXT: v_lshr_b64 v[34:35], v[9:10], 16
|
|
; SI-NEXT: v_lshr_b64 v[32:33], v[7:8], 16
|
|
; SI-NEXT: v_lshr_b64 v[27:28], v[3:4], 16
|
|
; SI-NEXT: .LBB59_3: ; %end
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v54
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v62
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v1
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v24
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v42
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 4, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v52
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v59
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v1
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v20
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v58
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v50
|
|
; SI-NEXT: v_and_b32_e32 v3, 0xffff, v57
|
|
; SI-NEXT: v_or_b32_e32 v1, v3, v1
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v16
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v46
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt vmcnt(6) expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v26
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v29
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v22
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v61
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v55
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v48
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v18
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v63
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v44
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v38
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v14
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v60
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v11
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v36
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v12
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v47
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v53
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v34
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v10
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v56
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v51
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v32
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v8
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v45
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v49
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v30
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v6
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v40
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v39
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v27
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v41
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v31
|
|
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v25
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v3
|
|
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
|
|
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
|
|
; SI-NEXT: s_waitcnt expcnt(0)
|
|
; SI-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v43
|
|
; SI-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x5c, v0
|
|
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
|
|
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v62, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
|
|
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
|
|
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; SI-NEXT: s_setpc_b64 s[30:31]
|
|
; SI-NEXT: .LBB59_4:
|
|
; SI-NEXT: s_branch .LBB59_2
|
|
;
|
|
; VI-LABEL: bitcast_v48f16_to_v48i16_scalar:
|
|
; VI: ; %bb.0:
|
|
; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v9
|
|
; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v8
|
|
; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v7
|
|
; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v6
|
|
; VI-NEXT: v_lshrrev_b32_e32 v19, 16, v5
|
|
; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v4
|
|
; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v3
|
|
; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v2
|
|
; VI-NEXT: v_lshrrev_b32_e32 v15, 16, v1
|
|
; VI-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; VI-NEXT: v_lshrrev_b32_e32 v14, 16, v0
|
|
; 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 v34, 0x200
|
|
; VI-NEXT: v_add_f16_e32 v30, s16, v34
|
|
; VI-NEXT: v_add_f16_e32 v55, s43, v34
|
|
; VI-NEXT: v_add_f16_e32 v31, s17, v34
|
|
; VI-NEXT: v_add_f16_e32 v54, s42, v34
|
|
; VI-NEXT: v_add_f16_e32 v32, s18, v34
|
|
; VI-NEXT: v_add_f16_e32 v53, s41, v34
|
|
; VI-NEXT: v_add_f16_e32 v33, s19, v34
|
|
; VI-NEXT: v_add_f16_e32 v52, s40, v34
|
|
; VI-NEXT: v_add_f16_e32 v24, s20, v34
|
|
; VI-NEXT: v_add_f16_e32 v51, s15, v34
|
|
; VI-NEXT: v_add_f16_e32 v25, s21, v34
|
|
; VI-NEXT: v_add_f16_e32 v50, s14, v34
|
|
; VI-NEXT: v_add_f16_e32 v26, s22, v34
|
|
; VI-NEXT: v_add_f16_e32 v49, s13, v34
|
|
; VI-NEXT: v_add_f16_e32 v27, s23, v34
|
|
; VI-NEXT: v_add_f16_e32 v48, s12, v34
|
|
; VI-NEXT: v_add_f16_e32 v28, s24, v34
|
|
; VI-NEXT: v_add_f16_e32 v39, s11, v34
|
|
; VI-NEXT: v_add_f16_e32 v29, s25, v34
|
|
; VI-NEXT: v_add_f16_e32 v38, s10, v34
|
|
; VI-NEXT: v_add_f16_e32 v10, s26, v34
|
|
; VI-NEXT: v_add_f16_e32 v37, s9, v34
|
|
; VI-NEXT: v_add_f16_e32 v11, s27, v34
|
|
; VI-NEXT: v_add_f16_e32 v36, s8, v34
|
|
; VI-NEXT: v_add_f16_e32 v12, s28, v34
|
|
; VI-NEXT: v_add_f16_e32 v35, s7, v34
|
|
; VI-NEXT: v_add_f16_e32 v13, s29, v34
|
|
; VI-NEXT: v_add_f16_e32 v34, s6, v34
|
|
; VI-NEXT: v_add_f16_e32 v0, 0x200, v0
|
|
; VI-NEXT: v_add_f16_e32 v14, 0x200, v14
|
|
; VI-NEXT: v_add_f16_e32 v1, 0x200, v1
|
|
; VI-NEXT: v_add_f16_e32 v15, 0x200, v15
|
|
; VI-NEXT: v_add_f16_e32 v2, 0x200, v2
|
|
; VI-NEXT: v_add_f16_e32 v16, 0x200, v16
|
|
; VI-NEXT: v_add_f16_e32 v3, 0x200, v3
|
|
; VI-NEXT: v_add_f16_e32 v17, 0x200, v17
|
|
; VI-NEXT: v_add_f16_e32 v4, 0x200, v4
|
|
; VI-NEXT: v_add_f16_e32 v18, 0x200, v18
|
|
; VI-NEXT: v_add_f16_e32 v5, 0x200, v5
|
|
; VI-NEXT: v_add_f16_e32 v19, 0x200, v19
|
|
; VI-NEXT: v_add_f16_e32 v6, 0x200, v6
|
|
; VI-NEXT: v_add_f16_e32 v20, 0x200, v20
|
|
; VI-NEXT: v_add_f16_e32 v7, 0x200, v7
|
|
; VI-NEXT: v_add_f16_e32 v21, 0x200, v21
|
|
; VI-NEXT: v_add_f16_e32 v8, 0x200, v8
|
|
; VI-NEXT: v_add_f16_e32 v22, 0x200, v22
|
|
; VI-NEXT: v_add_f16_e32 v9, 0x200, v9
|
|
; VI-NEXT: v_add_f16_e32 v23, 0x200, v23
|
|
; 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 v34, s6
|
|
; VI-NEXT: v_mov_b32_e32 v13, s29
|
|
; VI-NEXT: v_mov_b32_e32 v35, s7
|
|
; VI-NEXT: v_mov_b32_e32 v12, s28
|
|
; VI-NEXT: v_mov_b32_e32 v36, s8
|
|
; VI-NEXT: v_mov_b32_e32 v11, s27
|
|
; VI-NEXT: v_mov_b32_e32 v37, s9
|
|
; VI-NEXT: v_mov_b32_e32 v10, s26
|
|
; VI-NEXT: v_mov_b32_e32 v38, s10
|
|
; VI-NEXT: v_mov_b32_e32 v29, s25
|
|
; VI-NEXT: v_mov_b32_e32 v39, s11
|
|
; VI-NEXT: v_mov_b32_e32 v28, s24
|
|
; VI-NEXT: v_mov_b32_e32 v48, s12
|
|
; VI-NEXT: v_mov_b32_e32 v27, s23
|
|
; VI-NEXT: v_mov_b32_e32 v49, s13
|
|
; VI-NEXT: v_mov_b32_e32 v26, s22
|
|
; VI-NEXT: v_mov_b32_e32 v50, s14
|
|
; VI-NEXT: v_mov_b32_e32 v25, s21
|
|
; VI-NEXT: v_mov_b32_e32 v51, s15
|
|
; VI-NEXT: v_mov_b32_e32 v24, s20
|
|
; VI-NEXT: v_mov_b32_e32 v52, s40
|
|
; VI-NEXT: v_mov_b32_e32 v33, s19
|
|
; VI-NEXT: v_mov_b32_e32 v53, s41
|
|
; VI-NEXT: v_mov_b32_e32 v32, s18
|
|
; VI-NEXT: v_mov_b32_e32 v54, s42
|
|
; VI-NEXT: v_mov_b32_e32 v31, s17
|
|
; VI-NEXT: v_mov_b32_e32 v55, s43
|
|
; VI-NEXT: v_mov_b32_e32 v30, s16
|
|
; VI-NEXT: .LBB59_5: ; %end
|
|
; VI-NEXT: v_lshlrev_b32_e32 v14, 16, v14
|
|
; VI-NEXT: v_or_b32_sdwa v14, v0, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v15
|
|
; VI-NEXT: v_or_b32_sdwa v15, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v16
|
|
; VI-NEXT: v_or_b32_sdwa v16, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v17
|
|
; VI-NEXT: v_or_b32_sdwa v17, v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v18
|
|
; VI-NEXT: v_or_b32_sdwa v18, v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v19
|
|
; VI-NEXT: v_or_b32_sdwa v19, v5, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v20
|
|
; VI-NEXT: v_or_b32_sdwa v20, v6, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v21
|
|
; 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_or_b32_sdwa v21, v7, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v22
|
|
; VI-NEXT: v_or_b32_sdwa v30, v30, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v31, v31, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v32, v32, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v33, v33, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v24, v24, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v25, v25, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v26, v26, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v27, v27, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v28, v28, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v29, v29, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; 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_or_b32_sdwa v22, v8, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_lshlrev_b32_e32 v0, 16, v23
|
|
; VI-NEXT: v_or_b32_sdwa v10, v10, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v11, v11, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v12, v12, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v13, v13, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_or_b32_sdwa v23, v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
|
|
; VI-NEXT: v_mov_b32_e32 v0, v30
|
|
; VI-NEXT: v_mov_b32_e32 v1, v31
|
|
; VI-NEXT: v_mov_b32_e32 v2, v32
|
|
; VI-NEXT: v_mov_b32_e32 v3, v33
|
|
; VI-NEXT: v_mov_b32_e32 v4, v24
|
|
; VI-NEXT: v_mov_b32_e32 v5, v25
|
|
; VI-NEXT: v_mov_b32_e32 v6, v26
|
|
; VI-NEXT: v_mov_b32_e32 v7, v27
|
|
; VI-NEXT: v_mov_b32_e32 v8, v28
|
|
; VI-NEXT: v_mov_b32_e32 v9, v29
|
|
; VI-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: bitcast_v48f16_to_v48i16_scalar:
|
|
; GFX9: ; %bb.0:
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; 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: v_cmp_ne_u32_e32 vcc, 0, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v9
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v1
|
|
; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v0
|
|
; 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: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v9, v23, 16, v9
|
|
; GFX9-NEXT: s_movk_i32 s4, 0x200
|
|
; GFX9-NEXT: v_lshl_or_b32 v8, v22, 16, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v7, v21, 16, v7
|
|
; GFX9-NEXT: v_lshl_or_b32 v6, v20, 16, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v5, v19, 16, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v4, v18, 16, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v3, v17, 16, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v2, v16, 16, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v1, v15, 16, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v0, v14, 16, v0
|
|
; GFX9-NEXT: v_pk_add_f16 v9, v9, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v8, v8, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v7, v7, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v6, v6, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v5, v5, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v4, v4, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v3, v3, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v2, v2, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v1, v1, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_pk_add_f16 v0, v0, s4 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s29, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v14, 0x200
|
|
; GFX9-NEXT: v_pk_add_f16 v13, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s28, s42
|
|
; GFX9-NEXT: v_pk_add_f16 v12, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s27, s41
|
|
; GFX9-NEXT: v_pk_add_f16 v11, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s26, s40
|
|
; GFX9-NEXT: v_pk_add_f16 v10, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s25, s15
|
|
; GFX9-NEXT: v_pk_add_f16 v29, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s24, s14
|
|
; GFX9-NEXT: v_pk_add_f16 v28, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s23, s13
|
|
; GFX9-NEXT: v_pk_add_f16 v27, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s22, s12
|
|
; GFX9-NEXT: v_pk_add_f16 v26, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s21, s11
|
|
; GFX9-NEXT: v_pk_add_f16 v25, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s20, s10
|
|
; GFX9-NEXT: v_pk_add_f16 v24, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s19, s9
|
|
; GFX9-NEXT: v_pk_add_f16 v33, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s18, s8
|
|
; GFX9-NEXT: v_pk_add_f16 v32, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s17, s7
|
|
; GFX9-NEXT: v_pk_add_f16 v31, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s6
|
|
; GFX9-NEXT: v_pk_add_f16 v30, s4, v14 op_sel_hi:[1,0]
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v55, 16, v30
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v54, 16, v31
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v53, 16, v32
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v52, 16, v33
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v51, 16, v24
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v50, 16, v25
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v49, 16, v26
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v48, 16, v27
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v39, 16, v28
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v29
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v0
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v1
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v2
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v3
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v4
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v5
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v6
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v7
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v8
|
|
; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v9
|
|
; 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 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 v29, s25
|
|
; GFX9-NEXT: v_mov_b32_e32 v28, s24
|
|
; GFX9-NEXT: v_mov_b32_e32 v27, s23
|
|
; GFX9-NEXT: v_mov_b32_e32 v26, s22
|
|
; GFX9-NEXT: v_mov_b32_e32 v25, s21
|
|
; GFX9-NEXT: v_mov_b32_e32 v24, s20
|
|
; GFX9-NEXT: v_mov_b32_e32 v33, s19
|
|
; GFX9-NEXT: v_mov_b32_e32 v32, s18
|
|
; GFX9-NEXT: v_mov_b32_e32 v31, s17
|
|
; GFX9-NEXT: v_mov_b32_e32 v30, s16
|
|
; GFX9-NEXT: v_mov_b32_e32 v34, s43
|
|
; GFX9-NEXT: v_mov_b32_e32 v35, s42
|
|
; GFX9-NEXT: v_mov_b32_e32 v36, s41
|
|
; GFX9-NEXT: v_mov_b32_e32 v37, s40
|
|
; GFX9-NEXT: v_mov_b32_e32 v38, s15
|
|
; GFX9-NEXT: v_mov_b32_e32 v39, s14
|
|
; GFX9-NEXT: v_mov_b32_e32 v48, s13
|
|
; GFX9-NEXT: v_mov_b32_e32 v49, s12
|
|
; GFX9-NEXT: v_mov_b32_e32 v50, s11
|
|
; GFX9-NEXT: v_mov_b32_e32 v51, s10
|
|
; GFX9-NEXT: v_mov_b32_e32 v52, s9
|
|
; GFX9-NEXT: v_mov_b32_e32 v53, s8
|
|
; GFX9-NEXT: v_mov_b32_e32 v54, s7
|
|
; GFX9-NEXT: v_mov_b32_e32 v55, s6
|
|
; GFX9-NEXT: .LBB59_5: ; %end
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX9-NEXT: v_lshl_or_b32 v14, v14, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX9-NEXT: v_lshl_or_b32 v15, v15, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v2
|
|
; GFX9-NEXT: v_lshl_or_b32 v16, v16, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v3
|
|
; GFX9-NEXT: v_lshl_or_b32 v17, v17, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v4
|
|
; GFX9-NEXT: v_lshl_or_b32 v18, v18, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v5
|
|
; GFX9-NEXT: v_lshl_or_b32 v19, v19, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v6
|
|
; GFX9-NEXT: v_lshl_or_b32 v20, v20, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v7
|
|
; GFX9-NEXT: v_and_b32_e32 v30, 0xffff, v30
|
|
; GFX9-NEXT: v_and_b32_e32 v31, 0xffff, v31
|
|
; GFX9-NEXT: v_and_b32_e32 v32, 0xffff, v32
|
|
; GFX9-NEXT: v_and_b32_e32 v33, 0xffff, v33
|
|
; 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_and_b32_e32 v28, 0xffff, v28
|
|
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff, v29
|
|
; GFX9-NEXT: v_lshl_or_b32 v21, v21, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v8
|
|
; GFX9-NEXT: v_lshl_or_b32 v30, v55, 16, v30
|
|
; GFX9-NEXT: v_lshl_or_b32 v31, v54, 16, v31
|
|
; GFX9-NEXT: v_lshl_or_b32 v32, v53, 16, v32
|
|
; GFX9-NEXT: v_lshl_or_b32 v33, v52, 16, v33
|
|
; GFX9-NEXT: v_lshl_or_b32 v24, v51, 16, v24
|
|
; GFX9-NEXT: v_lshl_or_b32 v25, v50, 16, v25
|
|
; GFX9-NEXT: v_lshl_or_b32 v26, v49, 16, v26
|
|
; GFX9-NEXT: v_lshl_or_b32 v27, v48, 16, v27
|
|
; GFX9-NEXT: v_lshl_or_b32 v28, v39, 16, v28
|
|
; GFX9-NEXT: v_lshl_or_b32 v29, v38, 16, v29
|
|
; 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_lshl_or_b32 v22, v22, 16, v0
|
|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v9
|
|
; GFX9-NEXT: v_lshl_or_b32 v10, v37, 16, v10
|
|
; GFX9-NEXT: v_lshl_or_b32 v11, v36, 16, v11
|
|
; GFX9-NEXT: v_lshl_or_b32 v12, v35, 16, v12
|
|
; GFX9-NEXT: v_lshl_or_b32 v13, v34, 16, v13
|
|
; GFX9-NEXT: v_lshl_or_b32 v23, v23, 16, v0
|
|
; GFX9-NEXT: v_mov_b32_e32 v0, v30
|
|
; GFX9-NEXT: v_mov_b32_e32 v1, v31
|
|
; GFX9-NEXT: v_mov_b32_e32 v2, v32
|
|
; GFX9-NEXT: v_mov_b32_e32 v3, v33
|
|
; GFX9-NEXT: v_mov_b32_e32 v4, v24
|
|
; GFX9-NEXT: v_mov_b32_e32 v5, v25
|
|
; GFX9-NEXT: v_mov_b32_e32 v6, v26
|
|
; GFX9-NEXT: v_mov_b32_e32 v7, v27
|
|
; GFX9-NEXT: v_mov_b32_e32 v8, v28
|
|
; GFX9-NEXT: v_mov_b32_e32 v9, v29
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v48i16_scalar:
|
|
; GFX11-TRUE16: ; %bb.0:
|
|
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0
|
|
; 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 s13, s22, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s21, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s20, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s19, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s18, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s17, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s16, 16
|
|
; GFX11-TRUE16-NEXT: s_lshr_b32 s5, 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 s14, s0, 16
|
|
; GFX11-TRUE16-NEXT: s_mov_b32 s46, 0
|
|
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB59_3
|
|
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v23.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v24.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v21.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v24.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v20.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v24.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v19.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v24.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v18.h
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v24.h
|
|
; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s46
|
|
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB59_4
|
|
; GFX11-TRUE16-NEXT: .LBB59_2: ; %cmp.true
|
|
; 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 s13, s22, s13
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s21, s12
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s20, s11
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s19, s10
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s18, s9
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s17, s8
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s16, s6
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s5
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s7
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s14
|
|
; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s4
|
|
; 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, 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, s13 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s8 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 v55, 16, v0
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v1
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v2
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v3
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v4
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v5
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v6
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v8
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v9
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v10
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v12
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v13
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v14
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v15
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 16, v16
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v17
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v18
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v19
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v20
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v21
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v22
|
|
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v23
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB59_5
|
|
; GFX11-TRUE16-NEXT: .LBB59_3:
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25
|
|
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24
|
|
; GFX11-TRUE16-NEXT: s_branch .LBB59_2
|
|
; GFX11-TRUE16-NEXT: .LBB59_4:
|
|
; 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 v30, s45 :: v_dual_mov_b32 v31, s44
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v32, s43 :: v_dual_mov_b32 v33, s42
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v34, s41 :: v_dual_mov_b32 v35, s40
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v36, s15 :: v_dual_mov_b32 v37, s13
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v38, s12 :: v_dual_mov_b32 v39, s11
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v48, s10 :: v_dual_mov_b32 v49, s9
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v50, s8 :: v_dual_mov_b32 v51, s6
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v52, s5 :: v_dual_mov_b32 v53, s7
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s4 :: v_dual_mov_b32 v55, s14
|
|
; GFX11-TRUE16-NEXT: .LBB59_5: ; %end
|
|
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v55, v55 :: v_dual_mov_b32 v54, v54
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v53 :: v_dual_mov_b32 v52, v52
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v51, v51 :: v_dual_mov_b32 v50, v50
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v49, v49 :: v_dual_mov_b32 v48, v48
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v39, v39 :: v_dual_mov_b32 v38, v38
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v37, v37 :: v_dual_mov_b32 v36, v36
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v35, v35 :: v_dual_mov_b32 v34, v34
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v33, v33 :: v_dual_mov_b32 v32, v32
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, v31 :: v_dual_mov_b32 v30, v30
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v29 :: v_dual_mov_b32 v28, v28
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v27 :: v_dual_mov_b32 v26, v26
|
|
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v25 :: v_dual_mov_b32 v24, v24
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v55.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v54.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v53.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v52.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v51.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v50.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v49.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v48.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v39.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v38.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v37.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v36.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v35.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v34.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v33.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v32.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v31.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v30.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v29.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v28.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v27.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v26.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v25.l
|
|
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v24.l
|
|
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX11-FAKE16-LABEL: bitcast_v48f16_to_v48i16_scalar:
|
|
; 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, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v19, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v18, 16, v0
|
|
; 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 s7, s16, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s6, s3, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s8, s2, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s4, s1, 16
|
|
; GFX11-FAKE16-NEXT: s_lshr_b32 s5, s0, 16
|
|
; GFX11-FAKE16-NEXT: s_mov_b32 s46, 0
|
|
; GFX11-FAKE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
|
|
; 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, s46
|
|
; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB59_4
|
|
; GFX11-FAKE16-NEXT: .LBB59_2: ; %cmp.true
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; 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 v0, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v5, v23, 16, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v4, v22, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v3, v21, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v2, v20, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v1, v19, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v0, v18, 16, v0
|
|
; 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 s7, s16, s7
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s3, s3, s6
|
|
; GFX11-FAKE16-NEXT: s_pack_ll_b32_b16 s2, s2, s8
|
|
; 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 v5, 0x200, v5 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 v3, 0x200, v3 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 v1, 0x200, v1 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 v14, 0x200, s29 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v15, 0x200, s28 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v16, 0x200, s27 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v17, 0x200, s26 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v9, 0x200, s25 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v10, 0x200, s24 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v11, 0x200, s15 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v12, 0x200, s14 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v6, 0x200, s12 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v7, 0x200, s11 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v8, 0x200, s10 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v28, 0x200, s9 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v27, 0x200, s0 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v26, 0x200, s1 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v25, 0x200, s2 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v24, 0x200, s3 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_pk_add_f16 v29, 0x200, s7 op_sel_hi:[0,1]
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v25
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v24
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v28
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v6
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v10
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v9
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v31, 16, v15
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v30, 16, v14
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v18, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v19, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v20, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v21, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v22, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v23, 16, v5
|
|
; 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 v14, s29 :: v_dual_mov_b32 v15, s28
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v16, s27 :: v_dual_mov_b32 v17, s26
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v9, s25 :: v_dual_mov_b32 v10, s24
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v12, s22
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v13, s21 :: v_dual_mov_b32 v6, s20
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v8, s18
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v28, s17 :: v_dual_mov_b32 v29, s16
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v24, s3 :: v_dual_mov_b32 v25, s2
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v26, s1 :: v_dual_mov_b32 v27, s0
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v30, s45 :: v_dual_mov_b32 v31, s44
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v32, s43 :: v_dual_mov_b32 v33, s42
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v34, s41 :: v_dual_mov_b32 v35, s40
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v36, s15 :: v_dual_mov_b32 v37, s14
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v38, s13 :: v_dual_mov_b32 v39, s12
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v48, s11 :: v_dual_mov_b32 v49, s10
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v50, s9 :: v_dual_mov_b32 v51, s7
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v52, s6 :: v_dual_mov_b32 v53, s8
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v54, s4 :: v_dual_mov_b32 v55, s5
|
|
; GFX11-FAKE16-NEXT: .LBB59_5: ; %end
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v65, 0xffff, v24
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v24, v55, 16, v27
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v27, v52, 16, v65
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v64, 0xffff, v25
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v18, v18, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v25, v54, 16, v26
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v52, 0xffff, v28
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v22, v22, 16, v3
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, v27
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v26, v53, 16, v64
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v53, 0xffff, v6
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v6, v49, 16, v8
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v7, v48, 16, v7
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v8, v39, 16, v53
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v39, 0xffff, v10
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v48, 0xffff, v9
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v28, v51, 16, v29
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v29, v50, 16, v52
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v9, v38, 16, v13
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v10, v37, 16, v12
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v12, v35, 16, v39
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v13, v34, 16, v48
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v34, 0xffff, v15
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v35, 0xffff, v14
|
|
; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v5
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v11, v36, 16, v11
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v14, v33, 16, v17
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v15, v32, 16, v16
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v16, v31, 16, v34
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v17, v30, 16, v35
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v19, v19, 16, v0
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v20, v20, 16, v1
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v21, v21, 16, v2
|
|
; GFX11-FAKE16-NEXT: v_lshl_or_b32 v23, v23, 16, v4
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
|
|
; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, v26
|
|
; GFX11-FAKE16-NEXT: v_dual_mov_b32 v4, v28 :: v_dual_mov_b32 v5, 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 <48 x half> %a, splat (half 0xH0200)
|
|
%a2 = bitcast <48 x half> %a1 to <48 x i16>
|
|
br label %end
|
|
|
|
cmp.false:
|
|
%a3 = bitcast <48 x half> %a to <48 x i16>
|
|
br label %end
|
|
|
|
end:
|
|
%phi = phi <48 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
|
|
ret <48 x i16> %phi
|
|
}
|