This check was unnecessary/incorrect, it was already being done by the target
hook default implementation, and the one in the matcher was checking for a
completely different thing. This change:
1) Removes the check and updates affected tests which now do some more reassociations.
2) Modifies the AMDGPU hooks which were stubbed with "return true" to also do the oneuse
check. Not sure why I didn't do this the first time.
6959 lines
323 KiB
LLVM
6959 lines
323 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
|
|
; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdpal -mcpu=gfx600 < %s | FileCheck -check-prefixes=GFX67,GFX6,GFX67-SDAG,GFX6-SDAG %s
|
|
; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdpal -mcpu=gfx600 < %s | FileCheck -check-prefixes=GFX67,GFX6,GFX67-GISEL,GFX6-GISEL %s
|
|
|
|
; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX67,GFX7,GFX67-SDAG,GFX7-SDAG %s
|
|
; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX67,GFX7,GFX67-GISEL,GFX7-GISEL %s
|
|
|
|
; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdpal -mcpu=gfx803 < %s | FileCheck -check-prefixes=GFX8,GFX8-SDAG %s
|
|
; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdpal -mcpu=gfx803 < %s | FileCheck -check-prefixes=GFX8,GFX8-GISEL %s
|
|
|
|
; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX900,GFX9-SDAG,GFX900-SDAG %s
|
|
; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX900,GFX9-GISEL,GFX900-GISEL %s
|
|
|
|
; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefixes=GFX9,GFX90A,GFX9-SDAG,GFX90A-SDAG %s
|
|
; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefixes=GFX9,GFX90A,GFX9-GISEL,GFX90A-GISEL %s
|
|
|
|
; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdpal -mcpu=gfx1030 < %s | FileCheck -check-prefixes=GFX10,GFX10-SDAG %s
|
|
; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdpal -mcpu=gfx1030 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s
|
|
|
|
; Test for integer mad formation for patterns used in clpeak
|
|
|
|
define i32 @clpeak_imad_pat_i32(i32 %x, i32 %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_i32:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v1
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_i32:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_i32:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v0, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_i32:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: clpeak_imad_pat_i32:
|
|
; GFX900-SDAG: ; %bb.0: ; %entry
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v0, v2, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v1, v0, v[1:2]
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: clpeak_imad_pat_i32:
|
|
; GFX900-GISEL: ; %bb.0: ; %entry
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: clpeak_imad_pat_i32:
|
|
; GFX90A-SDAG: ; %bb.0: ; %entry
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v0, v2, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v0, v[2:3]
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: clpeak_imad_pat_i32:
|
|
; GFX90A-GISEL: ; %bb.0: ; %entry
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_i32:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[1:2], null, v0, v2, v[0:1]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v1, v0, v[1:2]
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_i32:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y18 = add i32 %x, 1
|
|
%add = mul i32 %y18, %y
|
|
%mul119 = add i32 %add, %y18
|
|
%add2 = mul i32 %mul119, %y
|
|
%add220 = add i32 %add, 1
|
|
%add422 = add i32 %add2, 1
|
|
%mul521 = mul i32 %add2, %add220
|
|
%add6 = mul i32 %mul521, %add422
|
|
ret i32 %add6
|
|
}
|
|
|
|
define signext i16 @clpeak_imad_pat_i16(i16 signext %x, i16 signext %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_i16:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v1, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_i16:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_i16:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_i16:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v3, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v3, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v3, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v1
|
|
; GFX8-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_imad_pat_i16:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_imad_pat_i16:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v3, 1, v1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v4, v3, v2
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v1, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v4
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v1, v3, v2, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_i16:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_i16:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v2, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v2, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%conv33 = add i16 %x, 1
|
|
%add = mul i16 %conv33, %y
|
|
%conv434 = add i16 %y, 1
|
|
%add8 = mul i16 %conv434, %add
|
|
%conv1035 = add i16 %add, 1
|
|
%add14 = mul i16 %conv1035, %add8
|
|
%conv1636 = add i16 %add8, 1
|
|
%add20 = mul i16 %add14, %conv1636
|
|
ret i16 %add20
|
|
}
|
|
|
|
define <2 x i16> @clpeak_imad_pat_v2i16(<2 x i16> %x, <2 x i16> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_v2i16:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v4, v3, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v6, v4, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v5, v2, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v5, v2, 1
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v5, 16, v6
|
|
; GFX67-SDAG-NEXT: s_mov_b32 s4, 0x10000
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v5, vcc, s4, v5
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, 1
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v2, 16, v1
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v5, 16, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 0x10000, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v4, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v5
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v2, 16, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v3, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_v2i16:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v5, v5, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v4, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v5, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v4, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v4, 16, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v2, 16, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_v2i16:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mov_b32_e32 v3, 1
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v2, 1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v3, 16, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v4, v0, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v3, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v5, v2, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v2, v1, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v0, v0, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v1, v2, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v0, v4, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v3, v1, v5, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v2, v0, v2
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v3, v1, v3
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_v2i16:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v3, 1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v2, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v2, v1, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v0, v3, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v6, v4, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v7, v5, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v2, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v4, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v5, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v6, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v7, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v2, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_imad_pat_v2i16:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v2, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_imad_pat_v2i16:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v2, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_v2i16:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v2, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_v2i16:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y18 = add <2 x i16> %x, <i16 1, i16 1>
|
|
%add = mul <2 x i16> %y18, %y
|
|
%mul119 = add <2 x i16> %add, %y18
|
|
%add2 = mul <2 x i16> %mul119, %y
|
|
%add220 = add <2 x i16> %add, <i16 1, i16 1>
|
|
%add422 = add <2 x i16> %add2, <i16 1, i16 1>
|
|
%mul521 = mul <2 x i16> %add2, %add220
|
|
%add6 = mul <2 x i16> %mul521, %add422
|
|
ret <2 x i16> %add6
|
|
}
|
|
|
|
define <3 x i16> @clpeak_imad_pat_v3i16(<3 x i16> %x, <3 x i16> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_v3i16:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v7, 16, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v7, 0, v7, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v7, v4, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v9, v7, v4
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v8, v3, v0
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v8, v8, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v6, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v1, 0, v1, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v6, v5, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v7, v0, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; GFX67-SDAG-NEXT: s_mov_b32 s4, 0x10000
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v8, vcc, s4, v8
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v3, 16, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v2, v5
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v6, v6, v5, 1
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v0, v3, v0
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v3, 0, v8, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v2, v5, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v7
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v7, 0xffff, v8
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, s4, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v5, v5, v7
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v3, v6
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v4, 0, v0, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v5, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v2, v3, v2
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_v3i16:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v6, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v6, v6, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v7, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v7, v7, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v8, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v6, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v8, v8, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v7, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v8, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v5, vcc, 1, v8
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v8, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v7
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v8
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_v3i16:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mov_b32_e32 v5, 1
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v4, 1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u16_sdwa v0, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v5, 16, v2
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v1, 1, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v6, v0, v5
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v5, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v7, v1, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v8, v4, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v4, v2, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v1, v3, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v0, v0, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v1, v1, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v2, v4, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v3, v0, v6, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v2, v8, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v3, v0, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v5, v1, v7, v1
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v4, v2, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v5, v1, v5
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_v3i16:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v6, 1
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v4, 16, v2
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v5, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u16_sdwa v0, v0, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v1, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v6, v5, v2, v5
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v7, v0, v4, v0
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v8, v1, v3, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v9, v6, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v10, v7, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v5, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v4, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v11, v8, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v1, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v6, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v7, v4, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v5, v9, v5
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v10, v0
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v8, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v11, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v5, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v0, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v1, v3
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_imad_pat_v3i16:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v1, -1
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v2, v4, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v3, v5, -1
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v4, v1, -1
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v5, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v5
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v4
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_imad_pat_v3i16:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v1, 1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v2, v4, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v3, v5, 1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v4, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v5, v1, 1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v4
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v5
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_v3i16:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v1, v1, -1
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v4, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v5, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v1, v4, v1
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v0, v5, v0
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v2, v5, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v3, v4, -1
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v4, v1, -1
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v5, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v5
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v4
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_v3i16:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v4, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v3, v5, 1
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v4, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v5, v1, 1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v4
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v5
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y48 = add <3 x i16> %x, <i16 1, i16 1, i16 1>
|
|
%add = mul <3 x i16> %y48, %y
|
|
%mul1249 = add <3 x i16> %add, %y48
|
|
%add15 = mul <3 x i16> %mul1249, %y
|
|
%add1550 = add <3 x i16> %add, <i16 1, i16 1, i16 1>
|
|
%add2452 = add <3 x i16> %add15, <i16 1, i16 1, i16 1>
|
|
%mul3051 = mul <3 x i16> %add15, %add1550
|
|
%add33 = mul <3 x i16> %mul3051, %add2452
|
|
ret <3 x i16> %add33
|
|
}
|
|
|
|
define <4 x i16> @clpeak_imad_pat_v4i16(<4 x i16> %x, <4 x i16> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_v4i16:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, 1, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v10, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v10, v7, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v9, 16, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v9, 0, v9, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v13, v10, v7
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v3, v7
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v7, v8, v6, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v11, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v8, v6, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v9, v5, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v8, 16, v13
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v12, v9, v5
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v11, v4, v0
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v8, v11, v4, 1
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v1, 0, v1, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v10, 16, v12
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v9, v0, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v5
|
|
; GFX67-SDAG-NEXT: s_mov_b32 s4, 0x10000
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v8, v10, v8
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v4, 1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v5, v2, v6
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v8, vcc, s4, v8
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v2, v6, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v4, 16, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v7, vcc, s4, v7
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v6, 16, v3
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v0, v4, v0
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v4, 0, v8, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v2, v6, v2
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v6, 16, v7
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v7
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, s4, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, s4, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v8, v9, v8
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v4, v5
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v3, v6
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v5, 0, v0, 16
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v6, 16, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v7, 0xffff, v8
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v7, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v5
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v2, v4, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v3, v6
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_v4i16:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v9, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v8, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v10, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v9, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v10, 16, v8
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v10, v10, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v11, 16, v9
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v8, v8, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v10, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v9, v9, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v11, v11, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v8, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v9, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, v11, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v2, 16, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v5, vcc, 1, v10
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v8
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v3, v3, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v9
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v11
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v6
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v6, 0xffff, v7
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v8, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v9, vcc, 1, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v7, 0xffff, v8
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v8, 0xffff, v9
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v8, 16, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v9, 16, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v8
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v3, v1, v9
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v6
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v7
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v2, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v7
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v3, v3, v5
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_v4i16:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mov_b32_e32 v6, 1
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v5, 1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u16_sdwa v0, v0, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v8, 16, v2
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v4, 1, v1
|
|
; GFX8-SDAG-NEXT: v_add_u16_sdwa v1, v1, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v6, 16, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v9, v0, v8
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v8, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v7, v1, v6
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v10, v5, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v11, v4, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v1, v6, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v4, v3, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v5, v5, v2, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v0, v0, v8
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v1, v1, v6
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v2, v5, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v3, v4, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v0, v9, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v5, v1, v7, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v7, v2, v10, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v4, v0, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v6, v3, v11, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v5, v1, v5
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v7, v2, v7
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v6, v3, v6
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_v4i16:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v5, 1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v4, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u16_sdwa v0, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v6, 1, v1
|
|
; GFX8-GISEL-NEXT: v_add_u16_sdwa v1, v1, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v2
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v7, 16, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v8, v4, v2, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v9, v0, v5, v0
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v10, v6, v3, v6
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v11, v1, v7, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v12, v8, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v13, v9, v5
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v4, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v5, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v14, v10, v3
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v15, v11, v7
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v6, v6, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v1, v7, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v8, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v9, v5, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v12, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v13, v0
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v10, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v7, v11, v7, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v6, v14, v6
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v15, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v4, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v0, v0, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v6, v3
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v1, v1, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_imad_pat_v4i16:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v1, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v2, v4, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v3, v5, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v4, v1, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v5, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v5
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v4
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_imad_pat_v4i16:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v1, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v2, v4, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v3, v5, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v4, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v5, v1, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v4
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v5
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_v4i16:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v1, v1, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v4, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v5, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v1, v4, v1
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v0, v5, v0
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v2, v5, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v3, v4, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v4, v1, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v5, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v5
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v4
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_v4i16:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v1, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v4, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v3, v5, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v4, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v5, v1, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v4
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v5
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y18 = add <4 x i16> %x, <i16 1, i16 1, i16 1, i16 1>
|
|
%add = mul <4 x i16> %y18, %y
|
|
%mul119 = add <4 x i16> %add, %y18
|
|
%add2 = mul <4 x i16> %mul119, %y
|
|
%add220 = add <4 x i16> %add, <i16 1, i16 1, i16 1, i16 1>
|
|
%add422 = add <4 x i16> %add2, <i16 1, i16 1, i16 1, i16 1>
|
|
%mul521 = mul <4 x i16> %add2, %add220
|
|
%add6 = mul <4 x i16> %mul521, %add422
|
|
ret <4 x i16> %add6
|
|
}
|
|
|
|
define zeroext i16 @clpeak_umad_pat_i16(i16 zeroext %x, i16 zeroext %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_umad_pat_i16:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v2, v0, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v3, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v1, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v3, v2, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_umad_pat_i16:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_umad_pat_i16:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_umad_pat_i16:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v3, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v3, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v3, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v1
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_umad_pat_i16:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_umad_pat_i16:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v3, 1, v1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v4, v3, v2
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v1, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v4
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v1, v3, v2, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_umad_pat_i16:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_umad_pat_i16:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v2, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v2, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%conv33 = add i16 %x, 1
|
|
%add = mul i16 %conv33, %y
|
|
%conv434 = add i16 %y, 1
|
|
%add8 = mul i16 %conv434, %add
|
|
%conv1035 = add i16 %add, 1
|
|
%add14 = mul i16 %conv1035, %add8
|
|
%conv1636 = add i16 %add8, 1
|
|
%add20 = mul i16 %add14, %conv1636
|
|
ret i16 %add20
|
|
}
|
|
|
|
define <2 x i16> @clpeak_umad_pat_v2i16(<2 x i16> %x, <2 x i16> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_umad_pat_v2i16:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v4, v3, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v6, v4, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v5, v2, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v5, v2, 1
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v5, 16, v6
|
|
; GFX67-SDAG-NEXT: s_mov_b32 s4, 0x10000
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v5, vcc, s4, v5
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, 1
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v2, 16, v1
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v5, 16, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 0x10000, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v4, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v5
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v2, 16, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v3, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_umad_pat_v2i16:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v5, v5, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v4, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v5, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v4, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v4, 16, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v2, 16, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_umad_pat_v2i16:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mov_b32_e32 v3, 1
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v2, 1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v3, 16, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v4, v0, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v3, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v5, v2, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v2, v1, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v0, v0, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v1, v2, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v0, v4, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v3, v1, v5, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v2, v0, v2
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v3, v1, v3
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_umad_pat_v2i16:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v3, 1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v2, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v2, v1, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v0, v3, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v6, v4, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v7, v5, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v2, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v4, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v5, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v6, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v7, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v2, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_umad_pat_v2i16:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v2, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_umad_pat_v2i16:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v2, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_umad_pat_v2i16:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v2, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_umad_pat_v2i16:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y18 = add <2 x i16> %x, <i16 1, i16 1>
|
|
%add = mul <2 x i16> %y18, %y
|
|
%mul119 = add <2 x i16> %add, %y18
|
|
%add2 = mul <2 x i16> %mul119, %y
|
|
%add220 = add <2 x i16> %add, <i16 1, i16 1>
|
|
%add422 = add <2 x i16> %add2, <i16 1, i16 1>
|
|
%mul521 = mul <2 x i16> %add2, %add220
|
|
%add6 = mul <2 x i16> %mul521, %add422
|
|
ret <2 x i16> %add6
|
|
}
|
|
|
|
define <3 x i16> @clpeak_umad_pat_v3i16(<3 x i16> %x, <3 x i16> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_umad_pat_v3i16:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v7, 16, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v7, 0, v7, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v7, v4, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v9, v7, v4
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v8, v3, v0
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v8, v8, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v6, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v1, 0, v1, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v6, v5, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v7, v0, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v8, v9, v8
|
|
; GFX67-SDAG-NEXT: s_mov_b32 s4, 0x10000
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v8, vcc, s4, v8
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v3, 16, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v2, v5
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v6, v6, v5, 1
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v0, v3, v0
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v3, 0, v8, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v2, v5, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v7
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v7, 0xffff, v8
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, s4, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v5, v5, v7
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v3, v6
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v4, 0, v0, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v5, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v2, v3, v2
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_umad_pat_v3i16:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v6, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v6, v6, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v7, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v7, v7, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v8, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v6, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v8, v8, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v7, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v8, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v5, vcc, 1, v8
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v8, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v7
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v8
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_umad_pat_v3i16:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mov_b32_e32 v5, 1
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v4, 1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u16_sdwa v0, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v5, 16, v2
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v1, 1, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v6, v0, v5
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v5, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v7, v1, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v8, v4, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v4, v2, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v1, v3, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v0, v0, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v1, v1, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v2, v4, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v3, v0, v6, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v2, v8, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v3, v0, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v5, v1, v7, v1
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v4, v2, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v5, v1, v5
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_umad_pat_v3i16:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v6, 1
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v4, 16, v2
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v5, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u16_sdwa v0, v0, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v1, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v6, v5, v2, v5
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v7, v0, v4, v0
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v8, v1, v3, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v9, v6, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v10, v7, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v5, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v4, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v11, v8, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v1, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v6, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v7, v4, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v5, v9, v5
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v10, v0
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v8, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v11, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v5, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v0, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v1, v3
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_umad_pat_v3i16:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v1, -1
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v2, v4, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v3, v5, -1
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v4, v1, -1
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v5, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v5
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v4
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_umad_pat_v3i16:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v1, 1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v2, v4, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v3, v5, 1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v4, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v5, v1, 1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v4
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v5
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_umad_pat_v3i16:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v1, v1, -1
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v4, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v5, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v1, v4, v1
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v0, v5, v0
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v2, v5, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v3, v4, -1
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v4, v1, -1
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v5, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v5
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v4
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_umad_pat_v3i16:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v4, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v3, v5, 1
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v4, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v5, v1, 1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v4
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v5
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y48 = add <3 x i16> %x, <i16 1, i16 1, i16 1>
|
|
%add = mul <3 x i16> %y48, %y
|
|
%mul1249 = add <3 x i16> %add, %y48
|
|
%add15 = mul <3 x i16> %mul1249, %y
|
|
%add1550 = add <3 x i16> %add, <i16 1, i16 1, i16 1>
|
|
%add2452 = add <3 x i16> %add15, <i16 1, i16 1, i16 1>
|
|
%mul3051 = mul <3 x i16> %add15, %add1550
|
|
%add33 = mul <3 x i16> %mul3051, %add2452
|
|
ret <3 x i16> %add33
|
|
}
|
|
|
|
define <4 x i16> @clpeak_umad_pat_v4i16(<4 x i16> %x, <4 x i16> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_umad_pat_v4i16:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, 1, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v10, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v10, v7, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v9, 16, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v9, 0, v9, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v13, v10, v7
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v3, v7
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v7, v8, v6, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v11, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v8, v6, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v9, v5, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v8, 16, v13
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v12, v9, v5
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v11, v4, v0
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v7, v8, v7
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v8, v11, v4, 1
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v1, 0, v1, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v10, 16, v12
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v9, v0, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v5
|
|
; GFX67-SDAG-NEXT: s_mov_b32 s4, 0x10000
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v8, v10, v8
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v4, 1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v5, v2, v6
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v8, vcc, s4, v8
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v2, v6, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v4, 16, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v7, vcc, s4, v7
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v6, 16, v3
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v0, v4, v0
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v4, 0, v8, 16
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v2, v6, v2
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v6, 16, v7
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v7
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, s4, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, s4, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v8, v9, v8
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v4, v5
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v3, v6
|
|
; GFX67-SDAG-NEXT: v_alignbit_b32 v5, 0, v0, 16
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v6, 16, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v7, 0xffff, v8
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v7, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v5
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v2, v4, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v3, v6
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_umad_pat_v4i16:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v9, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v8, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v9, 16, v9
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v10, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v8, v8, v9
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v9, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v10, 16, v10
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v9, v9, v10
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v10, 16, v8
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v8, 0xffff, v8
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v10, v10, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v11, 16, v9
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v8, v8, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v9, 0xffff, v9
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v10, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v9, v9, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v11, v11, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v8, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v9, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, v11, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v2, 16, v2
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v2, 16, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v5, vcc, 1, v10
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v8
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v3, v3, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v9
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v11
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v6
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v6, 0xffff, v7
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v5, v5, v6
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v8, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v9, vcc, 1, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v7, 16, v7
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v7, 0xffff, v8
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v8, 0xffff, v9
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v8, 16, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v9, 16, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v8
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v3, v1, v9
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v6
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v7
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v6, 0xffff, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v2, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v7
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v6
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v2, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v3, v3, v5
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_umad_pat_v4i16:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mov_b32_e32 v6, 1
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v5, 1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u16_sdwa v0, v0, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v8, 16, v2
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v4, 1, v1
|
|
; GFX8-SDAG-NEXT: v_add_u16_sdwa v1, v1, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v6, 16, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v9, v0, v8
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v8, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v7, v1, v6
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v10, v5, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v11, v4, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v1, v6, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v4, v3, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v5, v5, v2, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v0, v0, v8
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v1, v1, v6
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v2, v5, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v3, v4, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v0, v9, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v5, v1, v7, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v7, v2, v10, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v4, v0, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v6, v3, v11, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v5, v1, v5
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v7, v2, v7
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v6, v3, v6
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_umad_pat_v4i16:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v5, 1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v4, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u16_sdwa v0, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v6, 1, v1
|
|
; GFX8-GISEL-NEXT: v_add_u16_sdwa v1, v1, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v2
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v7, 16, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v8, v4, v2, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v9, v0, v5, v0
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v10, v6, v3, v6
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v11, v1, v7, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v12, v8, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v13, v9, v5
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v4, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v5, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v14, v10, v3
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v15, v11, v7
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v6, v6, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v1, v7, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v8, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v9, v5, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v12, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v13, v0
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v10, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v7, v11, v7, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v6, v14, v6
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v15, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v4, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v0, v0, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v6, v3
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v1, v1, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v1, v2, v1
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_umad_pat_v4i16:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v1, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v2, v4, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v3, v5, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v4, v1, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v5, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v5
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v4
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_umad_pat_v4i16:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v1, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v2, v4, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v3, v5, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v4, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v5, v1, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v4
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v5
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_umad_pat_v4i16:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v1, v1, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v4, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v5, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v1, v4, v1
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v0, v5, v0
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v2, v5, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v3, v4, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v4, v1, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v5, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v5
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v1, v4
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_umad_pat_v4i16:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v1, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v4, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v5, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v4, v0
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v5, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v4, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v3, v5, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v4, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v5, v1, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v4
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v1, v5
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y18 = add <4 x i16> %x, <i16 1, i16 1, i16 1, i16 1>
|
|
%add = mul <4 x i16> %y18, %y
|
|
%mul119 = add <4 x i16> %add, %y18
|
|
%add2 = mul <4 x i16> %mul119, %y
|
|
%add220 = add <4 x i16> %add, <i16 1, i16 1, i16 1, i16 1>
|
|
%add422 = add <4 x i16> %add2, <i16 1, i16 1, i16 1, i16 1>
|
|
%mul521 = mul <4 x i16> %add2, %add220
|
|
%add6 = mul <4 x i16> %mul521, %add422
|
|
ret <4 x i16> %add6
|
|
}
|
|
|
|
define <2 x i32> @clpeak_imad_pat_v2i32(<2 x i32> %x, <2 x i32> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_v2i32:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v4, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v5, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v0, v4
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v3, v1, v5
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, v2, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, v3, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v3
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_v2i32:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v4, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v5, v1
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v2, v1
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_v2i32:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, 1, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v4, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v5, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v0, v4
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v3, v1, v5
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v2, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, v3, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v0, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v3
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_v2i32:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, v4, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, v5, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, 1, v4
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v3, vcc, 1, v5
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v4, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v1, v3
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v2, v1
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: clpeak_imad_pat_v2i32:
|
|
; GFX900-SDAG: ; %bb.0: ; %entry
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, v4, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, v5, v1
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v1, v3
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v0, v4, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v5, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v0, v[3:4]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v4, v2, v[4:5]
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: clpeak_imad_pat_v2i32:
|
|
; GFX900-GISEL: ; %bb.0: ; %entry
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, v4, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, v5, v1
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v3, 1, v5
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v2, v0, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v3, v1, v3
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: clpeak_imad_pat_v2i32:
|
|
; GFX90A-SDAG: ; %bb.0: ; %entry
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v6, v0, v2
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v4, v1, v3
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, v4, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v6, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v1, v3
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v4, v[2:3]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v0, v6, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v0, v[6:7]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v2, v[4:5]
|
|
; GFX90A-SDAG-NEXT: v_mov_b32_e32 v1, v2
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: clpeak_imad_pat_v2i32:
|
|
; GFX90A-GISEL: ; %bb.0: ; %entry
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v4, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v5, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, 1, v5
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v4, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v5, 1, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_v2i32:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, v4, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, v5, v1
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v1, v3
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[3:4], null, v0, v4, v[0:1]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[4:5], null, v2, v5, v[2:3]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v3, v0, v[3:4]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[1:2], null, v4, v2, v[4:5]
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_v2i32:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, v4, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, v5, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, 1, v4
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v3, 1, v5
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v3, v1, v3
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y18 = add <2 x i32> %x, <i32 1, i32 1>
|
|
%add = mul <2 x i32> %y18, %y
|
|
%mul119 = add <2 x i32> %add, %y18
|
|
%add2 = mul <2 x i32> %mul119, %y
|
|
%add220 = add <2 x i32> %add, <i32 1, i32 1>
|
|
%add422 = add <2 x i32> %add2, <i32 1, i32 1>
|
|
%mul521 = mul <2 x i32> %add2, %add220
|
|
%add6 = mul <2 x i32> %mul521, %add422
|
|
ret <2 x i32> %add6
|
|
}
|
|
|
|
define <3 x i32> @clpeak_imad_pat_v3i32(<3 x i32> %x, <3 x i32> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_v3i32:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v6, v2, v5
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v7, v0, v3
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v8, v1, v4
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, v6, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v7, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v8, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v2, v5
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v3, v2, v6
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v4, v0, v7
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v5, v1, v8
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, v3, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v4, vcc, v4, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v5, vcc, v5, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v4, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v5, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v3, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v4
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v5
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, v2, v3
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_v3i32:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v6, v0, v3
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v7, v1, v4
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v8, v2, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v6, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v7, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v8, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v2, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v6
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v5, vcc, 1, v8
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v3, v2, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v6
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v1, v7
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v3, v2
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_v3i32:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, 1, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, 1, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v6, v2, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v7, v0, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v8, v1, v4
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v6, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v7, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v8, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v2, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v3, v2, v6
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v4, v0, v7
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v5, v1, v8
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, v3, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v4, vcc, v4, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v5, vcc, v5, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v4, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v5, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v3, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v0, v4
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v5
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v2, v3
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_v3i32:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v6, v0, v3
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v7, v1, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v8, v2, v5
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, v6, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, v7, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, v8, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v2, v5
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v3, vcc, 1, v6
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v4, vcc, 1, v7
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v5, vcc, 1, v8
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v6, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v7, vcc, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v3, v2, v5
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v6
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v1, v7
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v3, v2
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: clpeak_imad_pat_v3i32:
|
|
; GFX900-SDAG: ; %bb.0: ; %entry
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v2, 1, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v6, v2, v5
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v7, v0, v3
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v8, v1, v4
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v2, v6, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v5, v2, v5
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, v7, v0
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, v8, v1
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v1, v4
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v5, v6, v[5:6]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v0, v7, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v2, v8, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v0, v[6:7]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v7, v2, v[7:8]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v3, v5, v[3:4]
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: clpeak_imad_pat_v3i32:
|
|
; GFX900-GISEL: ; %bb.0: ; %entry
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v2, 1, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v6, v0, v3
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v7, v1, v4
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v8, v2, v5
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, v6, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, v7, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v2, v8, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v2, v2, v5
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v3, 1, v6
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v4, 1, v7
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v5, 1, v8
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v6, 1, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v7, 1, v1
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v3, v2, v5
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v2, 1, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v0, v6
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v1, v7
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v2, v3, v2
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: clpeak_imad_pat_v3i32:
|
|
; GFX90A-SDAG: ; %bb.0: ; %entry
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v2, 1, v2
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v10, v0, v3
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v8, v1, v4
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v6, v2, v5
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, v8, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v10, v0
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v7, v6, v2
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v1, v4
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v4, v7, v5
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v2, v8, v[2:3]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v0, v10, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v4, v6, v[4:5]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v0, v[10:11]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v8, v2, v[8:9]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v4, v[6:7]
|
|
; GFX90A-SDAG-NEXT: v_mov_b32_e32 v1, v8
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: clpeak_imad_pat_v3i32:
|
|
; GFX90A-GISEL: ; %bb.0: ; %entry
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, 1, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v6, v0, v3
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v7, v1, v4
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v8, v2, v5
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v6, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v7, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, v8, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v2, v2, v5
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, 1, v6
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v4, 1, v7
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v5, 1, v8
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v6, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v7, 1, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v8, 1, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v2, v2, v5
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v6
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v1, v7
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v2, v2, v8
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_v3i32:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v2, 1, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v6, v0, v3
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v7, v1, v4
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v8, v2, v5
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, v6, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, v7, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v9, v8, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v1, v4
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v3, v9, v5
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[4:5], null, v0, v6, v[0:1]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[5:6], null, v2, v7, v[2:3]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[6:7], null, v3, v8, v[3:4]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v4, v0, v[4:5]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[1:2], null, v5, v2, v[5:6]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[2:3], null, v6, v3, v[6:7]
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_v3i32:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, 1, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v6, v0, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v7, v1, v4
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v8, v2, v5
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, v6, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, v7, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, v8, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v0, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v2, v5
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v3, 1, v6
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v4, 1, v7
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v5, 1, v8
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v3, v0, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v4, v1, v4
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v5, v2, v5
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, 1, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v3, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v4, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v5, v2
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y48 = add <3 x i32> %x, <i32 1, i32 1, i32 1>
|
|
%add = mul <3 x i32> %y48, %y
|
|
%mul1249 = add <3 x i32> %add, %y48
|
|
%add15 = mul <3 x i32> %mul1249, %y
|
|
%add1550 = add <3 x i32> %add, <i32 1, i32 1, i32 1>
|
|
%add2452 = add <3 x i32> %add15, <i32 1, i32 1, i32 1>
|
|
%mul3051 = mul <3 x i32> %add15, %add1550
|
|
%add33 = mul <3 x i32> %mul3051, %add2452
|
|
ret <3 x i32> %add33
|
|
}
|
|
|
|
define <4 x i32> @clpeak_imad_pat_v4i32(<4 x i32> %x, <4 x i32> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_v4i32:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, 1, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v8, v0, v4
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v9, v3, v7
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v10, v1, v5
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v11, v2, v6
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v8, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, v9, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, v11, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v10, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v3, v3, v7
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v2, v6
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v4, v3, v9
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v5, v2, v11
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v6, v0, v8
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v7, v1, v10
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v4, vcc, v4, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v5, vcc, v5, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v6, vcc, v6, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v7, vcc, v7, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v6, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v7, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v5, v2
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v3, v4, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v6
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v7
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, v2, v5
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, v3, v4
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_v4i32:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v3
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v8, v0, v4
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v9, v1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v10, v2, v6
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v11, v3, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v8, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v9, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v10, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, v11, v3
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v2, v6
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v3, v3, v7
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v8
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v5, vcc, 1, v9
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v10
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v11
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v8, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v9, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v10, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v11, vcc, 1, v3
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v2, v6
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v3, v3, v7
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v8
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v1, v9
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v2, v10
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v3, v3, v11
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_v4i32:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, 1, v3
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, 1, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, 1, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v8, v0, v4
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v9, v3, v7
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v10, v1, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v11, v2, v6
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v8, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, v9, v3
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v11, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v10, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v3, v3, v7
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v2, v6
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v4, v3, v9
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v5, v2, v11
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v6, v0, v8
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v7, v1, v10
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v4, vcc, v4, v3
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v5, vcc, v5, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v6, vcc, v6, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v7, vcc, v7, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v6, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v7, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v5, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v3, v4, v3
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v0, v6
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v7
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v2, v5
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, v3, v4
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_v4i32:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v3, vcc, 1, v3
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v8, v0, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v9, v1, v5
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v10, v2, v6
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v11, v3, v7
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, v8, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, v9, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, v10, v2
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v3, vcc, v11, v3
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v2, v6
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v3, v3, v7
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v4, vcc, 1, v8
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v5, vcc, 1, v9
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v6, vcc, 1, v10
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v7, vcc, 1, v11
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v8, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v9, vcc, 1, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v10, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v11, vcc, 1, v3
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v2, v6
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v3, v3, v7
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v8
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v1, v9
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v2, v10
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v3, v3, v11
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: clpeak_imad_pat_v4i32:
|
|
; GFX900-SDAG: ; %bb.0: ; %entry
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v3, 1, v3
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v2, 1, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v8, v3, v7
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v11, v0, v4
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v9, v2, v6
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v10, v1, v5
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v3, v8, v3
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, v11, v0
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v12, v9, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v4, v3, v7
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v3, v12, v6
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, v10, v1
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v1, v5
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v4, v8, v[4:5]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v3, v9, v[3:4]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v0, v11, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v2, v10, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v7, v0, v[7:8]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v8, v2, v[8:9]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v3, v[6:7]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v5, v4, v[5:6]
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_imad_pat_v4i32:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v2, 1, v2
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v3, 1, v3
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v8, v0, v4
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v9, v1, v5
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v10, v2, v6
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v11, v3, v7
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v0, v8, v0
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v1, v9, v1
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v2, v10, v2
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v3, v11, v3
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v2, v2, v6
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v3, v3, v7
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v4, 1, v8
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v5, 1, v9
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v6, 1, v10
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v7, 1, v11
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v8, 1, v0
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v9, 1, v1
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v10, 1, v2
|
|
; GFX9-GISEL-NEXT: v_add_u32_e32 v11, 1, v3
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v2, v2, v6
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v3, v3, v7
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v0, v0, v8
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v1, v1, v9
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v2, v2, v10
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u32 v3, v3, v11
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: clpeak_imad_pat_v4i32:
|
|
; GFX90A-SDAG: ; %bb.0: ; %entry
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v3, 1, v3
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v2, 1, v2
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v14, v0, v4
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v12, v1, v5
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v10, v2, v6
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v8, v3, v7
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v3, v8, v3
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v9, v10, v2
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, v12, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v14, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v1, v5
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v4, v9, v6
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v6, v3, v7
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v6, v8, v[6:7]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v4, v10, v[4:5]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v2, v12, v[2:3]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v0, v14, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v14, v0, v[14:15]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v12, v2, v[12:13]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v10, v4, v[10:11]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v8, v6, v[8:9]
|
|
; GFX90A-SDAG-NEXT: v_mov_b32_e32 v1, v12
|
|
; GFX90A-SDAG-NEXT: v_mov_b32_e32 v3, v4
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_v4i32:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v2, 1, v2
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v3, 1, v3
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v8, v0, v4
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v9, v1, v5
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v10, v2, v6
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v11, v3, v7
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, v8, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, v9, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v12, v10, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v4, v11, v3
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v1, v5
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v3, v12, v6
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v4, v4, v7
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[5:6], null, v0, v8, v[0:1]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[6:7], null, v2, v9, v[2:3]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[7:8], null, v3, v10, v[3:4]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[8:9], null, v4, v11, v[4:5]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v5, v0, v[5:6]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[1:2], null, v6, v2, v[6:7]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[2:3], null, v7, v3, v[7:8]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[3:4], null, v8, v4, v[8:9]
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_v4i32:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, 1, v2
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v3, 1, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v8, v0, v4
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v9, v1, v5
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v10, v2, v6
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v11, v3, v7
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, v8, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, v9, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, v10, v2
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v3, v11, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v2, v6
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v3, v3, v7
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v4, 1, v8
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v5, 1, v9
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v6, 1, v10
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v7, 1, v11
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v8, 1, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v4, v1, v5
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v5, v2, v6
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v6, v3, v7
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, 1, v2
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v3, 1, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v0, v8
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v4, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v5, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v3, v6, v3
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y18 = add <4 x i32> %x, <i32 1, i32 1, i32 1, i32 1>
|
|
%add = mul <4 x i32> %y18, %y
|
|
%mul119 = add <4 x i32> %add, %y18
|
|
%add2 = mul <4 x i32> %mul119, %y
|
|
%add220 = add <4 x i32> %add, <i32 1, i32 1, i32 1, i32 1>
|
|
%add422 = add <4 x i32> %add2, <i32 1, i32 1, i32 1, i32 1>
|
|
%mul521 = mul <4 x i32> %add2, %add220
|
|
%add6 = mul <4 x i32> %mul521, %add422
|
|
ret <4 x i32> %add6
|
|
}
|
|
|
|
define i32 @clpeak_imad_pat_i24(i32 %x, i32 %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_i24:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 24
|
|
; GFX67-SDAG-NEXT: v_bfe_i32 v1, v1, 0, 24
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v1
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_i24:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 24
|
|
; GFX67-GISEL-NEXT: v_bfe_i32 v1, v1, 0, 24
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_i24:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 24
|
|
; GFX8-SDAG-NEXT: v_bfe_i32 v1, v1, 0, 24
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v0, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_i24:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 24
|
|
; GFX8-GISEL-NEXT: v_bfe_i32 v1, v1, 0, 24
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: clpeak_imad_pat_i24:
|
|
; GFX900-SDAG: ; %bb.0: ; %entry
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 24
|
|
; GFX900-SDAG-NEXT: v_bfe_i32 v1, v1, 0, 24
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v0, v2, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v1, v0, v[1:2]
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: clpeak_imad_pat_i24:
|
|
; GFX900-GISEL: ; %bb.0: ; %entry
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 24
|
|
; GFX900-GISEL-NEXT: v_bfe_i32 v1, v1, 0, 24
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: clpeak_imad_pat_i24:
|
|
; GFX90A-SDAG: ; %bb.0: ; %entry
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 24
|
|
; GFX90A-SDAG-NEXT: v_bfe_i32 v1, v1, 0, 24
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v0, v2, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v0, v[2:3]
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: clpeak_imad_pat_i24:
|
|
; GFX90A-GISEL: ; %bb.0: ; %entry
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 24
|
|
; GFX90A-GISEL-NEXT: v_bfe_i32 v1, v1, 0, 24
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_i24:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 24
|
|
; GFX10-SDAG-NEXT: v_bfe_i32 v1, v1, 0, 24
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[1:2], null, v0, v2, v[0:1]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v1, v0, v[1:2]
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_i24:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 24
|
|
; GFX10-GISEL-NEXT: v_bfe_i32 v1, v1, 0, 24
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%shl = shl i32 %x, 8
|
|
%shr = ashr exact i32 %shl, 8
|
|
%shl1 = shl i32 %y, 8
|
|
%shr2 = ashr exact i32 %shl1, 8
|
|
%shr222 = add nsw i32 %shr, 1
|
|
%add = mul i32 %shr2, %shr222
|
|
%mul323 = add i32 %add, %shr222
|
|
%add4 = mul i32 %mul323, %shr2
|
|
%add424 = add i32 %add, 1
|
|
%add626 = add i32 %add4, 1
|
|
%mul725 = mul i32 %add4, %add424
|
|
%add8 = mul i32 %mul725, %add626
|
|
ret i32 %add8
|
|
}
|
|
|
|
define i32 @clpeak_imad_pat_u24(i32 %x, i32 %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_u24:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffffff, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v1
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_u24:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffffff, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_u24:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_and_b32_e32 v0, 0xffffff, v0
|
|
; GFX8-SDAG-NEXT: v_and_b32_e32 v1, 0xffffff, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v0, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_u24:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_and_b32_e32 v0, 0xffffff, v0
|
|
; GFX8-GISEL-NEXT: v_and_b32_e32 v1, 0xffffff, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: clpeak_imad_pat_u24:
|
|
; GFX900-SDAG: ; %bb.0: ; %entry
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_and_b32_e32 v0, 0xffffff, v0
|
|
; GFX900-SDAG-NEXT: v_and_b32_e32 v1, 0xffffff, v1
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v0, v2, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v1, v0, v[1:2]
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: clpeak_imad_pat_u24:
|
|
; GFX900-GISEL: ; %bb.0: ; %entry
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_and_b32_e32 v0, 0xffffff, v0
|
|
; GFX900-GISEL-NEXT: v_and_b32_e32 v1, 0xffffff, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: clpeak_imad_pat_u24:
|
|
; GFX90A-SDAG: ; %bb.0: ; %entry
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_and_b32_e32 v0, 0xffffff, v0
|
|
; GFX90A-SDAG-NEXT: v_and_b32_e32 v1, 0xffffff, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v0, v2, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v0, v[2:3]
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: clpeak_imad_pat_u24:
|
|
; GFX90A-GISEL: ; %bb.0: ; %entry
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_and_b32_e32 v0, 0xffffff, v0
|
|
; GFX90A-GISEL-NEXT: v_and_b32_e32 v1, 0xffffff, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_u24:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_and_b32_e32 v0, 0xffffff, v0
|
|
; GFX10-SDAG-NEXT: v_and_b32_e32 v1, 0xffffff, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[1:2], null, v0, v2, v[0:1]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v1, v0, v[1:2]
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_u24:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_and_b32_e32 v0, 0xffffff, v0
|
|
; GFX10-GISEL-NEXT: v_and_b32_e32 v1, 0xffffff, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%shl = and i32 %x, 16777215
|
|
%shl1 = and i32 %y, 16777215
|
|
%shl122 = add nuw nsw i32 %shl, 1
|
|
%add = mul i32 %shl1, %shl122
|
|
%mul323 = add i32 %add, %shl122
|
|
%add4 = mul i32 %mul323, %shl1
|
|
%add424 = add i32 %add, 1
|
|
%add626 = add i32 %add4, 1
|
|
%mul725 = mul i32 %add4, %add424
|
|
%add8 = mul i32 %mul725, %add626
|
|
ret i32 %add8
|
|
}
|
|
|
|
define signext i8 @clpeak_imad_pat_i8(i8 signext %x, i8 signext %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_i8:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xff, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v1, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 8
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_i8:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 8
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_i8:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 8
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_i8:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v3, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v3, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v3, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v1
|
|
; GFX8-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 8
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_imad_pat_i8:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 8
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_imad_pat_i8:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v3, 1, v1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v4, v3, v2
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v1, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v4
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v1, v3, v2, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 8
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_i8:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 8
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_i8:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v2, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v2, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 8
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%conv33 = add i8 %x, 1
|
|
%add = mul i8 %conv33, %y
|
|
%conv434 = add i8 %y, 1
|
|
%add8 = mul i8 %conv434, %add
|
|
%conv1035 = add i8 %add, 1
|
|
%add14 = mul i8 %conv1035, %add8
|
|
%conv1636 = add i8 %add8, 1
|
|
%add20 = mul i8 %add14, %conv1636
|
|
ret i8 %add20
|
|
}
|
|
|
|
define <2 x i8> @clpeak_imad_pat_v2i8(<2 x i8> %x, <2 x i8> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_v2i8:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xff, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v5, v3, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v6, v5, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v4, v2, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v4, v2, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xff, v3
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v4, 8, v6
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v5, v0, v2
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v3, v4, v3
|
|
; GFX67-SDAG-NEXT: s_movk_i32 s4, 0x100
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, 1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, s4, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v2, 8, v1
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; GFX67-SDAG-NEXT: v_bfe_u32 v2, v3, 8, 8
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xff, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 0x100, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v4, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-SDAG-NEXT: v_bfe_u32 v2, v0, 8, 8
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v3, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_v2i8:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v4, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xff, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v5, v5, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v4, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v5, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xff, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v5, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xff, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xff, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_v2i8:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v1, 1, v1
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v4, v1, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v1, v3, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v5, v0, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v2, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v1, v1, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v0, v0, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v3, v1, v4, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v0, v5, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v3, v1, v3
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b16_e32 v3, 8, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v2, v0, v2
|
|
; GFX8-SDAG-NEXT: v_or_b32_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_v2i8:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v1, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v0, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v1, v3, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v6, v4, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v7, v5, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v1, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v4, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v5, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v6, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v7, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v1, v3
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_imad_pat_v2i8:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_add_u16_e32 v1, 1, v1
|
|
; GFX9-SDAG-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX9-SDAG-NEXT: v_mul_lo_u16_e32 v4, v1, v3
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v1, v3, v1
|
|
; GFX9-SDAG-NEXT: v_mul_lo_u16_e32 v5, v0, v2
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v0, v2, v0
|
|
; GFX9-SDAG-NEXT: v_mul_lo_u16_e32 v1, v1, v3
|
|
; GFX9-SDAG-NEXT: v_mul_lo_u16_e32 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v3, v1, v4, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v2, v0, v5, v0
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v3, v1, v3
|
|
; GFX9-SDAG-NEXT: v_lshlrev_b16_e32 v3, 8, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v2, v0, v2
|
|
; GFX9-SDAG-NEXT: v_or_b32_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
|
|
; GFX9-SDAG-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_imad_pat_v2i8:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v1, 1, v1
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v4, v0, v2, v0
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v5, v1, v3, v1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v6, v4, v2
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v7, v5, v3
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v2, 1
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v1, v1, v3, 1
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v2, v4, v2, 1
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v3, v5, v3, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v6, v0
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v1, v7, v1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v1, v1, v3
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_v2i8:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v4, v1, v3, v1
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u16 v1, v1, v3
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v5, v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u16 v3, v4, v3
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u16 v2, v5, v2
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v3, v1, v3
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v2, v0, v2
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v1, v3, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_lshlrev_b16 v2, 8, v1
|
|
; GFX10-SDAG-NEXT: v_and_b32_e32 v1, 0xff, v1
|
|
; GFX10-SDAG-NEXT: v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_v2i8:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v4, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v5, v1, v3
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v4, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v5, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v2, v4, 1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v3, v5, 1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v4, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v5, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v4
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v1, v5
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y18 = add <2 x i8> %x, <i8 1, i8 1>
|
|
%add = mul <2 x i8> %y18, %y
|
|
%mul119 = add <2 x i8> %add, %y18
|
|
%add2 = mul <2 x i8> %mul119, %y
|
|
%add220 = add <2 x i8> %add, <i8 1, i8 1>
|
|
%add422 = add <2 x i8> %add2, <i8 1, i8 1>
|
|
%mul521 = mul <2 x i8> %add2, %add220
|
|
%add6 = mul <2 x i8> %mul521, %add422
|
|
ret <2 x i8> %add6
|
|
}
|
|
|
|
define i64 @clpeak_imad_pat_i64(i64 %x, i64 %y) {
|
|
; GFX6-SDAG-LABEL: clpeak_imad_pat_i64:
|
|
; GFX6-SDAG: ; %bb.0: ; %entry
|
|
; GFX6-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v4, v0, v3
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v5, v0, v2
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v6, v1, v2
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v4, vcc, v5, v4
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v5, v0, v2
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v4, vcc, v4, v6
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, v5, v0
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v3, v0, v3
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v6, v0, v2
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v1, vcc, v4, v1, vcc
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v1, v1, v2
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v2, v0, v4
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v4, v0, v5
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v3, vcc, v6, v3
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v1, vcc, v3, v1
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v3, v1, v5
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v2, vcc, v4, v2
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v4, v0, v5
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v2, vcc, v2, v3
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v3, vcc, v4, v0
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v2, vcc, v2, v1, vcc
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v4, v3, v0
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v5, v2, v0
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v0, v3, v0
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v1, vcc, v4, v1
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v5
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v3
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v1, vcc, v1, v2, vcc
|
|
; GFX6-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX6-GISEL-LABEL: clpeak_imad_pat_i64:
|
|
; GFX6-GISEL: ; %bb.0: ; %entry
|
|
; GFX6-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v4, v1, v2
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v5, v0, v3
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v7, v0, v2
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v6, v0, v2
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v5
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v7
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v6, v0
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v1, vcc, v4, v1, vcc
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v1, v1, v2
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v3, v0, v3
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v5, v0, v2
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v0, v0, v2
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v3
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v1, v0
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v6
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v2, vcc, 0, v4, vcc
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v4, vcc, 0, v0, vcc
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v2, v5, v2
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v6, v5, v1
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v1, v5, v1
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v2
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v1
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v1, v0, v3
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v2, v6, v4
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v0, v6, v3
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v3, v6, v3
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v2
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v3
|
|
; GFX6-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX7-SDAG-LABEL: clpeak_imad_pat_i64:
|
|
; GFX7-SDAG: ; %bb.0: ; %entry
|
|
; GFX7-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v4, vcc, 1, v0
|
|
; GFX7-SDAG-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v6, v4, v3
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v2, 0
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v7, v5, v2
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v6
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v7
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v4, vcc, v0, v4
|
|
; GFX7-SDAG-NEXT: v_addc_u32_e32 v5, vcc, v1, v5, vcc
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v6, v4, v3
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v4, v2, 0
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v2, v5, v2
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v4, vcc, v4, v6
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v4, vcc, v4, v2
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v2, v4, v0
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v3, v0, v[3:4]
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v0, v3, v1
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v1, vcc, v2, v6
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v6, vcc, v0, v1
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v2, v6, v3
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v3, v[5:6]
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v3, v5, v4
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v1, vcc, v2, v1
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v1, vcc, v3, v1
|
|
; GFX7-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX7-GISEL-LABEL: clpeak_imad_pat_i64:
|
|
; GFX7-GISEL: ; %bb.0: ; %entry
|
|
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v0
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v2, 0
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v3, v[1:2]
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v8, vcc, v0, v6
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v2, v[4:5]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v2, 0
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v9, vcc, v4, v7, vcc
|
|
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v6
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v8, v3, v[1:2]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v9, v2, v[6:7]
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v0
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v4, vcc
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v5, v6, 0
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v5
|
|
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v4
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v1, v[0:1]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v7, 0
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v8, vcc, 0, v2, vcc
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v6, v[4:5]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v8, v[1:2]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v4, v7, v[1:2]
|
|
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_i64:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v4, vcc, 1, v0
|
|
; GFX8-SDAG-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v6, v4, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v2, 0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v7, v5, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v6
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v7
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v4, vcc, v0, v4
|
|
; GFX8-SDAG-NEXT: v_addc_u32_e32 v5, vcc, v1, v5, vcc
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v6, v4, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v4, v2, 0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v5, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v4, vcc, v4, v6
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v4, vcc, v4, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v4, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v3, v0, v[3:4]
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v3, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v2, v6
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v6, vcc, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v6, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v3, v[5:6]
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v3, v5, v4
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v2, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v3, v1
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_i64:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v6, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v2, 0
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v3, v[1:2]
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v8, vcc, v0, v6
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v2, v[4:5]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v2, 0
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v9, vcc, v4, v7, vcc
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v6
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v8, v3, v[1:2]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v9, v2, v[6:7]
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v6, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v4, vcc
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v5, v6, 0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v7, vcc, 1, v5
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v1, v[0:1]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v7, 0
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v8, vcc, 0, v2, vcc
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v6, v[4:5]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v8, v[1:2]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v4, v7, v[1:2]
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: clpeak_imad_pat_i64:
|
|
; GFX900-SDAG: ; %bb.0: ; %entry
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_add_co_u32_e32 v4, vcc, 1, v0
|
|
; GFX900-SDAG-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v6, v5, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v7, v4, v3
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v2, 0
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v6, v1, v7, v6
|
|
; GFX900-SDAG-NEXT: v_add_co_u32_e32 v1, vcc, v0, v4
|
|
; GFX900-SDAG-NEXT: v_addc_co_u32_e32 v4, vcc, v6, v5, vcc
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v4, v4, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v3, v1, v3
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v1, v2, 0
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v2, v2, v3, v4
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v5, v2, v0
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v1, v0, v[1:2]
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v1, v6
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v3, v2
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v4, v5, v4, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v5, v4, v1
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v1, v[3:4]
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v1, v5, v1, v2
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: clpeak_imad_pat_i64:
|
|
; GFX900-GISEL: ; %bb.0: ; %entry
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v0
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v1, vcc
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v2, 0
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v3, v[1:2]
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, v0, v6
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v2, v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v2, 0
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v9, vcc, v4, v7, vcc
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v6
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v8, v3, v[1:2]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v9, v2, v[6:7]
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v0
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v4, vcc
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v5, v6, 0
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v7, vcc, 1, v5
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, v4
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v1, v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v7, 0
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v8, vcc, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v6, v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v3, v8, v[1:2]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v4, v7, v[1:2]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: clpeak_imad_pat_i64:
|
|
; GFX90A-SDAG: ; %bb.0: ; %entry
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_co_u32_e32 v0, vcc, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v6, v1, v2
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v7, v0, v3
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v0, v2, 0
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v0, v2, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v1, v6, v1, v7
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v5, v5, v7, v6
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v3, v0, v3
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v6, v1, v2
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v0, v2, 0
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v1, v1, v3, v6
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v6, v1, v4
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v0, v4, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v4, v0, v5
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v3, v6, v3, v4
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v4, v2, v1
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v5, v3, v0
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v0, v[2:3]
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v1, v5, v1, v4
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: clpeak_imad_pat_i64:
|
|
; GFX90A-GISEL: ; %bb.0: ; %entry
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v1, vcc
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v3, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v2, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v2, v[4:5]
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v1, v4
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, v0, v6
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v8, vcc, v1, v7, vcc
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v2, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v6, v3, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v2, v[6:7]
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v0, vcc, 0, v1, vcc
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v5, v5, v2
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v7, vcc, 1, v4
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v0, 0
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v8, vcc, 0, v5, vcc
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v6, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v6, v[0:1]
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v4, v3, v0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v7, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v2, v8, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v7, v[2:3]
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_i64:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_add_co_u32 v4, vcc_lo, v0, 1
|
|
; GFX10-SDAG-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v7, v4, v3
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v4, v2, 0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v6, v5, v2
|
|
; GFX10-SDAG-NEXT: v_add_co_u32 v4, vcc_lo, v0, v4
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v1, v1, v7, v6
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v6, v4, v3
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[3:4], null, v4, v2, 0
|
|
; GFX10-SDAG-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, v1, v5, vcc_lo
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v5, v2
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v4, v4, v6, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v3, v1
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v5, v4, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v3, v0, v[3:4]
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v1, v5, v1, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v0, v4
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v4, v1, v3
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v0, v3, v[0:1]
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v1, v4, v1, v2
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_i64:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v4, vcc_lo, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[0:1], null, v4, v2, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v6, v4, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v7, v5, v2
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v1, v1, v6, v7
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v6, vcc_lo, v0, v4
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v7, vcc_lo, v1, v5, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[4:5], null, v6, v2, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v3, v6, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v7, v2
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v0, vcc_lo, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v4, v1
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v5, v5, v3, v2
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[2:3], null, v4, v0, 0
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v4, vcc_lo, v4, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v5, v0
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, 0, v5, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v3, v3, v1, v0
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[0:1], null, v2, v4, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v2, v5
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v3, v3, v4
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v1, v1, v2, v3
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y18 = add i64 %x, 1
|
|
%add = mul i64 %y18, %y
|
|
%mul119 = add i64 %add, %y18
|
|
%add2 = mul i64 %mul119, %y
|
|
%add220 = add i64 %add, 1
|
|
%add422 = add i64 %add2, 1
|
|
%mul521 = mul i64 %add2, %add220
|
|
%add6 = mul i64 %mul521, %add422
|
|
ret i64 %add6
|
|
}
|
|
|
|
define <2 x i64> @clpeak_imad_pat_v2i64(<2 x i64> %x, <2 x i64> %y) {
|
|
; GFX6-SDAG-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX6-SDAG: ; %bb.0: ; %entry
|
|
; GFX6-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v8, v0, v5
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v9, v0, v4
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v10, v1, v4
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v8, vcc, v9, v8
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v8, vcc, v8, v10
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v9, v2, v7
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v10, v2, v6
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v12, v3, v6
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v11, v0, v4
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v9, vcc, v10, v9
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v9, vcc, v9, v12
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v10, v2, v6
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, v11, v0
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v1, vcc, v8, v1, vcc
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v5, v0, v5
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v12, v0, v4
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v2, vcc, v10, v2
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v3, vcc, v9, v3, vcc
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v5, vcc, v12, v5
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v1, vcc, v5, v1
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v5, v2, v7
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v7, v2, v6
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v2, v2, v6
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v3, v3, v6
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v4, vcc, v7, v5
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v5, v2, v9
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v6, v2, v10
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v3, vcc, v4, v3
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v4, v3, v10
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v5, vcc, v6, v5
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v6, v0, v11
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v4, vcc, v5, v4
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v5, v0, v8
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v8, v1, v11
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v7, v2, v10
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v5, vcc, v6, v5
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v6, v0, v11
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v5, vcc, v5, v8
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v6, vcc, v6, v0
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v5, vcc, v5, v1, vcc
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v1, v6, v1
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v8, v6, v0
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v9, v5, v0
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v0, v6, v0
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v7, vcc, v7, v2
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v4, vcc, v4, v3, vcc
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v1, vcc, v8, v1
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v9
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v3, v7, v3
|
|
; GFX6-SDAG-NEXT: v_mul_hi_u32 v8, v7, v2
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v6
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v1, vcc, v1, v5, vcc
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v5, v4, v2
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v2, v7, v2
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v3, vcc, v8, v3
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v3, vcc, v3, v5
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v2, vcc, v2, v7
|
|
; GFX6-SDAG-NEXT: v_addc_u32_e32 v3, vcc, v3, v4, vcc
|
|
; GFX6-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX6-GISEL-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX6-GISEL: ; %bb.0: ; %entry
|
|
; GFX6-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v8, v1, v4
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v9, v0, v5
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v11, v0, v4
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v2
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v8, vcc, v8, v9
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v8, vcc, v8, v11
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v9, v3, v6
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v11, v2, v7
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v13, v2, v6
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v10, v0, v4
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v12, v2, v6
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v9, vcc, v9, v11
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v9, vcc, v9, v13
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v10, v0
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v1, vcc, v8, v1, vcc
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v1, v1, v4
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v5, v0, v5
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v11, v0, v4
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v0, v0, v4
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v2, vcc, v12, v2
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v3, vcc, v9, v3, vcc
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v5
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v1, v0
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v1, v3, v6
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v3, v2, v7
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v4, v2, v6
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v2, v2, v6
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v3
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v2
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v10
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v3, vcc, 0, v8, vcc
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v5, vcc, 1, v12
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v6, vcc, 0, v9, vcc
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v7, vcc, 1, v11
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v8, vcc, 0, v0, vcc
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v3, v11, v3
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v12, v11, v2
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v2, v11, v2
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v9, vcc, 1, v4
|
|
; GFX6-GISEL-NEXT: v_addc_u32_e32 v10, vcc, 0, v1, vcc
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v3
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v2
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v2, v4, v6
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v3, v4, v5
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v4, v4, v5
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v5, v12, v7
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v2
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v2, vcc, v1, v4
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v1, v0, v7
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v4, v12, v8
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v0, v12, v7
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v4
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v5
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v4, v2, v9
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v5, v3, v10
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v2, v3, v9
|
|
; GFX6-GISEL-NEXT: v_mul_hi_u32 v3, v3, v9
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v5
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v3, vcc, v4, v3
|
|
; GFX6-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX7-SDAG-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX7-SDAG: ; %bb.0: ; %entry
|
|
; GFX7-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v8, vcc, 1, v0
|
|
; GFX7-SDAG-NEXT: v_addc_u32_e32 v9, vcc, 0, v1, vcc
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v10, v8, v5
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v8, v4, 0
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v11, vcc, 1, v2
|
|
; GFX7-SDAG-NEXT: v_addc_u32_e32 v12, vcc, 0, v3, vcc
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v3, vcc, v1, v10
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v10, v9, v4
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v13, v11, v7
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v11, v6, 0
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v14, vcc, v3, v10
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v3, v12, v6
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v2, vcc, v2, v13
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v13, vcc, v2, v3
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v2, vcc, v0, v8
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v5, v2, v5
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v2, v4, 0
|
|
; GFX7-SDAG-NEXT: v_addc_u32_e32 v9, vcc, v14, v9, vcc
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v8, vcc, v1, v11
|
|
; GFX7-SDAG-NEXT: v_addc_u32_e32 v10, vcc, v13, v12, vcc
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v3, vcc, v3, v5
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v5, v8, v7
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v8, v6, 0
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v9, v9, v4
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v4, v10, v6
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v5, vcc, v8, v5
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v3, vcc, v3, v9
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v8, vcc, v5, v4
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v6, v8, v1
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v1, v[7:8]
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v1, v3, v0
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v2, v0, v[2:3]
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v5, vcc, v6, v5
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v6, v7, v13
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v0, v2, v14
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v10
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v5, vcc, v6, v5
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v10, vcc, v0, v1
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v11, v10, v2
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v9, v2, v[9:10]
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v6, v9, v3
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v9, v5, v7
|
|
; GFX7-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v7, v[4:5]
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v4, v4, v8
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v1, vcc, v11, v1
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v3, vcc, v9, v3
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v1, vcc, v6, v1
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v3, vcc, v4, v3
|
|
; GFX7-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX7-GISEL-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX7-GISEL: ; %bb.0: ; %entry
|
|
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v12, vcc, 1, v0
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v13, vcc, 0, v1, vcc
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v14, vcc, 1, v2
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v4, 0
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v3, vcc
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v14, v6, 0
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v5, v[1:2]
|
|
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v3
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v14, v7, v[1:2]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v4, v[8:9]
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v3, vcc, v0, v12
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v15, v6, v[10:11]
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v18, vcc, v8, v13, vcc
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v4, 0
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v16, vcc, v2, v14
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v16, v6, 0
|
|
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v11
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v19, vcc, v9, v15, vcc
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v3, v5, v[1:2]
|
|
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v13
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v16, v7, v[1:2]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v18, v4, v[14:15]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v19, v6, v[16:17]
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v11, vcc, 1, v0
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v10, v11, 0
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v8, vcc
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v13, vcc, 1, v2
|
|
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v6
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v1, v[0:1]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v12, v13, 0
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v2, vcc, 0, v9, vcc
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v14, vcc, 1, v10
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v3, v11, v[0:1]
|
|
; GFX7-GISEL-NEXT: v_mov_b32_e32 v0, v7
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v3, vcc
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v12, v2, v[0:1]
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v16, vcc, 1, v12
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v14, 0
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v4, v13, v[2:3]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v16, 0
|
|
; GFX7-GISEL-NEXT: v_addc_u32_e32 v17, vcc, 0, v4, vcc
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v15, v[1:2]
|
|
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v3
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v6, v17, v[1:2]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v14, v[4:5]
|
|
; GFX7-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v16, v[10:11]
|
|
; GFX7-GISEL-NEXT: v_mov_b32_e32 v1, v5
|
|
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v8, vcc, 1, v0
|
|
; GFX8-SDAG-NEXT: v_addc_u32_e32 v9, vcc, 0, v1, vcc
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v10, v8, v5
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v8, v4, 0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v11, vcc, 1, v2
|
|
; GFX8-SDAG-NEXT: v_addc_u32_e32 v12, vcc, 0, v3, vcc
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, v1, v10
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v10, v9, v4
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v13, v11, v7
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v11, v6, 0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v14, vcc, v3, v10
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v3, v12, v6
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v2, v13
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v13, vcc, v2, v3
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v0, v8
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v5, v2, v5
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v2, v4, 0
|
|
; GFX8-SDAG-NEXT: v_addc_u32_e32 v9, vcc, v14, v9, vcc
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v8, vcc, v1, v11
|
|
; GFX8-SDAG-NEXT: v_addc_u32_e32 v10, vcc, v13, v12, vcc
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, v3, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v5, v8, v7
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v8, v6, 0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v9, v9, v4
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v4, v10, v6
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v5, vcc, v8, v5
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, v3, v9
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v8, vcc, v5, v4
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v6, v8, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v1, v[7:8]
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v3, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v2, v0, v[2:3]
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v5, vcc, v6, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v6, v7, v13
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v2, v14
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v10
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v5, vcc, v6, v5
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v10, vcc, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v11, v10, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v9, v2, v[9:10]
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v6, v9, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v9, v5, v7
|
|
; GFX8-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v7, v[4:5]
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v4, v4, v8
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v11, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, v9, v3
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v6, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, v4, v3
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v12, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v13, vcc, 0, v1, vcc
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v14, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v4, 0
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v3, vcc
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v14, v6, 0
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v5, v[1:2]
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v14, v7, v[1:2]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v4, v[8:9]
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v3, vcc, v0, v12
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v15, v6, v[10:11]
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v18, vcc, v8, v13, vcc
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v4, 0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v16, vcc, v2, v14
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v16, v6, 0
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v11
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v19, vcc, v9, v15, vcc
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v3, v5, v[1:2]
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v13
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v16, v7, v[1:2]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v18, v4, v[14:15]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v19, v6, v[16:17]
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v11, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v10, v11, 0
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v8, vcc
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v13, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v6
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v1, v[0:1]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v12, v13, 0
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v2, vcc, 0, v9, vcc
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v14, vcc, 1, v10
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v3, v11, v[0:1]
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v7
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v3, vcc
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v12, v2, v[0:1]
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v16, vcc, 1, v12
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v14, 0
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v4, v13, v[2:3]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v16, 0
|
|
; GFX8-GISEL-NEXT: v_addc_u32_e32 v17, vcc, 0, v4, vcc
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v15, v[1:2]
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v6, v17, v[1:2]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v14, v[4:5]
|
|
; GFX8-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v16, v[10:11]
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, v5
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX900-SDAG: ; %bb.0: ; %entry
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_add_co_u32_e32 v8, vcc, 1, v2
|
|
; GFX900-SDAG-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v3, vcc
|
|
; GFX900-SDAG-NEXT: v_add_co_u32_e32 v10, vcc, 1, v0
|
|
; GFX900-SDAG-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v1, vcc
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v12, v11, v4
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v13, v10, v5
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v4, 0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v14, v9, v6
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v15, v8, v7
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v6, 0
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v12, v1, v13, v12
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v1, v3, v15, v14
|
|
; GFX900-SDAG-NEXT: v_add_co_u32_e32 v8, vcc, v2, v8
|
|
; GFX900-SDAG-NEXT: v_addc_co_u32_e32 v9, vcc, v1, v9, vcc
|
|
; GFX900-SDAG-NEXT: v_add_co_u32_e32 v3, vcc, v0, v10
|
|
; GFX900-SDAG-NEXT: v_addc_co_u32_e32 v10, vcc, v12, v11, vcc
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v10, v10, v4
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v11, v3, v5
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v3, v4, 0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v9, v9, v6
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v7, v8, v7
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v6, 0
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v4, v4, v11, v10
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v10, v4, v0
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v6, v6, v7, v9
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v9, v6, v2
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v5, v2, v[5:6]
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v5, v1
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v0, v[3:4]
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v11, v3, v12
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v8, v9, v8, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v4, v0, v4
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v6, v7, v6
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v1, v10, v1, v11
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v9, v1, v3
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v0, v3, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v10, v8, v5
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v7, v5, v[7:8]
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v1, v9, v1, v4
|
|
; GFX900-SDAG-NEXT: v_add3_u32 v3, v10, v3, v6
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX900-GISEL: ; %bb.0: ; %entry
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 1, v0
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v1, vcc
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 1, v2
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v12, v4, 0
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v14, v6, 0
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v5, v[1:2]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v3
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v14, v7, v[1:2]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v4, v[8:9]
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v3, vcc, v0, v12
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v15, v6, v[10:11]
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v18, vcc, v8, v13, vcc
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v3, v4, 0
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v16, vcc, v2, v14
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[12:13], s[4:5], v16, v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v11
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v19, vcc, v9, v15, vcc
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[14:15], s[4:5], v3, v5, v[1:2]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v13
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[16:17], s[4:5], v16, v7, v[1:2]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v18, v4, v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v19, v6, v[16:17]
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v11, vcc, 1, v0
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v10, v11, 0
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v8, vcc
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v13, vcc, 1, v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, v6
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v1, v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v12, v13, 0
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v2, vcc, 0, v9, vcc
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 1, v10
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v3, v11, v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, v7
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v12, v2, v[0:1]
|
|
; GFX900-GISEL-NEXT: v_add_co_u32_e32 v16, vcc, 1, v12
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v14, 0
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[9:10], s[4:5], v4, v13, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v16, 0
|
|
; GFX900-GISEL-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v4, vcc
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v5, v15, v[1:2]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v3
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v6, v17, v[1:2]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v8, v14, v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v16, v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, v5
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX90A-SDAG: ; %bb.0: ; %entry
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_co_u32_e32 v2, vcc, 1, v2
|
|
; GFX90A-SDAG-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
|
|
; GFX90A-SDAG-NEXT: v_add_co_u32_e32 v0, vcc, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v12, v1, v4
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v13, v0, v5
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v0, v4, 0
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v0, v4, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v14, v3, v6
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v15, v2, v7
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v2, v6, 0
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v2, v6, v[2:3]
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v1, v12, v1, v13
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v9, v9, v13, v12
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v3, v14, v3, v15
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v5, v0, v5
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v12, v1, v4
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v0, v4, 0
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v1, v1, v5, v12
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v4, v2, v7
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v5, v3, v6
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v2, v6, 0
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v11, v11, v15, v14
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v3, v3, v4, v5
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v6, v3, v10
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v10, v[2:3]
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v7, v2, v11
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v5, v6, v5, v7
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v10, v1, v8
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v0, v8, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v8, v0, v9
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v7, v10, v7, v8
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v8, v6, v1
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v9, v7, v0
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v0, v[6:7]
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v6, v4, v3
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v7, v5, v2
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v2, v[4:5]
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v1, v9, v1, v8
|
|
; GFX90A-SDAG-NEXT: v_add3_u32 v3, v7, v3, v6
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX90A-GISEL: ; %bb.0: ; %entry
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v10, vcc, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v1, vcc
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 1, v2
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v3, vcc
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v10, v5, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v10, v4, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v11, v4, v[2:3]
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v7, 0
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v10, vcc, v0, v10
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v12, v6, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v6, v[8:9]
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v14, vcc, v1, v11, vcc
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, v3, v8
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, v2, v12
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v10, v4, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v10, v5, 0
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, v3, v13, vcc
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v14, v4, v[10:11]
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[10:11], s[4:5], v12, v7, 0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v9, v9, v4
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v12, v6, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v13, v6, v[10:11]
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v5, v5, v6
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v0, vcc, 0, v1, vcc
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v10, vcc, 1, v2
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v11, vcc, 0, v3, vcc
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v8, v0, 0
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 1, v8
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v6, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v9, v6, v[0:1]
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v9, vcc
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v8, v3, v0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v4, v11, 0
|
|
; GFX90A-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 1, v4
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v4, v10, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v5, v10, v[0:1]
|
|
; GFX90A-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v5, vcc
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v7, v7, v0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v12, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v2, v13, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v8, v12, v[2:3]
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v6, v15, 0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v1, v2
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v6, v14, 0
|
|
; GFX90A-GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v7, v14, v[4:5]
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, v3, v4
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_add_co_u32 v8, vcc_lo, v0, 1
|
|
; GFX10-SDAG-NEXT: v_add_co_ci_u32_e32 v9, vcc_lo, 0, v1, vcc_lo
|
|
; GFX10-SDAG-NEXT: v_add_co_u32 v10, vcc_lo, v2, 1
|
|
; GFX10-SDAG-NEXT: v_add_co_ci_u32_e32 v11, vcc_lo, 0, v3, vcc_lo
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v12, v9, v4
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v13, v8, v5
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v8, v4, 0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v14, v11, v6
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v15, v10, v7
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[2:3], null, v10, v6, 0
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v12, v1, v13, v12
|
|
; GFX10-SDAG-NEXT: v_add_co_u32 v1, vcc_lo, v0, v8
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v13, v3, v15, v14
|
|
; GFX10-SDAG-NEXT: v_add_co_ci_u32_e32 v3, vcc_lo, v12, v9, vcc_lo
|
|
; GFX10-SDAG-NEXT: v_add_co_u32 v8, vcc_lo, v2, v10
|
|
; GFX10-SDAG-NEXT: v_add_co_ci_u32_e32 v9, vcc_lo, v13, v11, vcc_lo
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v10, v3, v4
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v11, v1, v5
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v7, v8, v7
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v9, v9, v6
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[5:6], null, v8, v6, 0
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[3:4], null, v1, v4, 0
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v6, v6, v7, v9
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v4, v4, v11, v10
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v11, v3, v12
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v9, v6, v2
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[7:8], null, v5, v2, v[5:6]
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v10, v4, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v3, v0, v[3:4]
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v5, v13
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v6, v7, v6
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v1, v10, v1, v11
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v4, v0, v4
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v8, v9, v8, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v9, v1, v3
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v0, v3, v[0:1]
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v10, v8, v5
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[2:3], null, v7, v5, v[7:8]
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v1, v9, v1, v4
|
|
; GFX10-SDAG-NEXT: v_add3_u32 v3, v10, v3, v6
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_v2i64:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v8, vcc_lo, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v9, vcc_lo, 0, v1, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v10, vcc_lo, v2, 1
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v11, vcc_lo, 0, v3, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[0:1], null, v8, v4, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v12, v8, v5
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v13, v9, v4
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[2:3], null, v10, v6, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v14, v10, v7
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v15, v11, v6
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v1, v1, v12, v13
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v12, v3, v14, v15
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v3, vcc_lo, v0, v8
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v13, vcc_lo, v1, v9, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v10, vcc_lo, v2, v10
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v11, vcc_lo, v12, v11, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[8:9], null, v3, v4, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v5, v3, v5
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v13, v13, v4
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[3:4], null, v10, v6, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v7, v10, v7
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v6, v11, v6
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v0, vcc_lo, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v2, vcc_lo, v2, 1
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v9, v9, v5, v13
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v10, v4, v7, v6
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v11, vcc_lo, 0, v12, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[4:5], null, v8, v0, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v8, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v9, v0
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[6:7], null, v3, v2, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v11, v3, v11
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v10, v2
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v8, vcc_lo, v8, 1
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v9, vcc_lo, 0, v9, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_add_co_u32 v12, vcc_lo, v3, 1
|
|
; GFX10-GISEL-NEXT: v_add_co_ci_u32_e32 v10, vcc_lo, 0, v10, vcc_lo
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v3, v5, v1, v0
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v5, v7, v11, v2
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[0:1], null, v4, v8, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v4, v4, v9
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v7, v3, v8
|
|
; GFX10-GISEL-NEXT: v_mad_u64_u32 v[2:3], null, v6, v12, 0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v6, v6, v10
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v5, v5, v12
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v1, v1, v4, v7
|
|
; GFX10-GISEL-NEXT: v_add3_u32 v3, v3, v6, v5
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y18 = add <2 x i64> %x, <i64 1, i64 1>
|
|
%add = mul <2 x i64> %y18, %y
|
|
%mul119 = add <2 x i64> %add, %y18
|
|
%add2 = mul <2 x i64> %mul119, %y
|
|
%add220 = add <2 x i64> %add, <i64 1, i64 1>
|
|
%add422 = add <2 x i64> %add2, <i64 1, i64 1>
|
|
%mul521 = mul <2 x i64> %add2, %add220
|
|
%add6 = mul <2 x i64> %mul521, %add422
|
|
ret <2 x i64> %add6
|
|
}
|
|
|
|
define i32 @v_multi_use_mul_chain_add_other_use_all(i32 %arg, i32 %arg1, i32 %arg2, ptr addrspace(1) %ptr) {
|
|
; GFX6-SDAG-LABEL: v_multi_use_mul_chain_add_other_use_all:
|
|
; GFX6-SDAG: ; %bb.0: ; %bb
|
|
; GFX6-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX6-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX6-SDAG-NEXT: s_mov_b32 s7, 0xf000
|
|
; GFX6-SDAG-NEXT: s_mov_b32 s4, s6
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX6-SDAG-NEXT: s_mov_b32 s5, s6
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v5, v0, v1
|
|
; GFX6-SDAG-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
|
|
; GFX6-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX6-SDAG-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
|
|
; GFX6-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX6-SDAG-NEXT: buffer_store_dword v5, v[3:4], s[4:7], 0 addr64
|
|
; GFX6-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(1)
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, v5, v1
|
|
; GFX6-SDAG-NEXT: s_waitcnt expcnt(0)
|
|
; GFX6-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX6-GISEL-LABEL: v_multi_use_mul_chain_add_other_use_all:
|
|
; GFX6-GISEL: ; %bb.0: ; %bb
|
|
; GFX6-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX6-GISEL-NEXT: s_mov_b32 s6, 0
|
|
; GFX6-GISEL-NEXT: s_mov_b32 s7, 0xf000
|
|
; GFX6-GISEL-NEXT: s_mov_b64 s[4:5], 0
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v5, v0, v1
|
|
; GFX6-GISEL-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
|
|
; GFX6-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX6-GISEL-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
|
|
; GFX6-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX6-GISEL-NEXT: buffer_store_dword v5, v[3:4], s[4:7], 0 addr64
|
|
; GFX6-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(1)
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v5, v1
|
|
; GFX6-GISEL-NEXT: s_waitcnt expcnt(0)
|
|
; GFX6-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX7-SDAG-LABEL: v_multi_use_mul_chain_add_other_use_all:
|
|
; GFX7-SDAG: ; %bb.0: ; %bb
|
|
; GFX7-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX7-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX7-SDAG-NEXT: s_mov_b32 s7, 0xf000
|
|
; GFX7-SDAG-NEXT: s_mov_b32 s4, s6
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX7-SDAG-NEXT: s_mov_b32 s5, s6
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v5, v0, v1
|
|
; GFX7-SDAG-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
|
|
; GFX7-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX7-SDAG-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
|
|
; GFX7-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX7-SDAG-NEXT: buffer_store_dword v5, v[3:4], s[4:7], 0 addr64
|
|
; GFX7-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v0, vcc, v5, v1
|
|
; GFX7-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX7-GISEL-LABEL: v_multi_use_mul_chain_add_other_use_all:
|
|
; GFX7-GISEL: ; %bb.0: ; %bb
|
|
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX7-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX7-GISEL-NEXT: s_mov_b32 s6, 0
|
|
; GFX7-GISEL-NEXT: s_mov_b32 s7, 0xf000
|
|
; GFX7-GISEL-NEXT: s_mov_b64 s[4:5], 0
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX7-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX7-GISEL-NEXT: v_mul_lo_u32 v5, v0, v1
|
|
; GFX7-GISEL-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
|
|
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX7-GISEL-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
|
|
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX7-GISEL-NEXT: buffer_store_dword v5, v[3:4], s[4:7], 0 addr64
|
|
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v0, vcc, v5, v1
|
|
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-LABEL: v_multi_use_mul_chain_add_other_use_all:
|
|
; GFX8: ; %bb.0: ; %bb
|
|
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
; GFX8-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 1, v2
|
|
; GFX8-NEXT: v_mul_lo_u32 v5, v0, v1
|
|
; GFX8-NEXT: flat_store_dword v[3:4], v2
|
|
; GFX8-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX8-NEXT: flat_store_dword v[3:4], v0
|
|
; GFX8-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX8-NEXT: flat_store_dword v[3:4], v5
|
|
; GFX8-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v5, v1
|
|
; GFX8-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-LABEL: v_multi_use_mul_chain_add_other_use_all:
|
|
; GFX900: ; %bb.0: ; %bb
|
|
; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX900-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX900-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX900-NEXT: v_mul_lo_u32 v5, v0, v1
|
|
; GFX900-NEXT: global_store_dword v[3:4], v2, off
|
|
; GFX900-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX900-NEXT: global_store_dword v[3:4], v0, off
|
|
; GFX900-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX900-NEXT: global_store_dword v[3:4], v5, off
|
|
; GFX900-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX900-NEXT: v_add_u32_e32 v0, v5, v1
|
|
; GFX900-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: v_multi_use_mul_chain_add_other_use_all:
|
|
; GFX90A-SDAG: ; %bb.0: ; %bb
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX90A-SDAG-NEXT: v_mov_b32_e32 v5, v4
|
|
; GFX90A-SDAG-NEXT: v_mov_b32_e32 v4, v3
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX90A-SDAG-NEXT: global_store_dword v[4:5], v2, off
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX90A-SDAG-NEXT: global_store_dword v[4:5], v0, off
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-SDAG-NEXT: global_store_dword v[4:5], v0, off
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v0, v1
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: v_multi_use_mul_chain_add_other_use_all:
|
|
; GFX90A-GISEL: ; %bb.0: ; %bb
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_mov_b32_e32 v2, v3
|
|
; GFX90A-GISEL-NEXT: v_mov_b32_e32 v3, v4
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v4, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v4, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v4
|
|
; GFX90A-GISEL-NEXT: global_store_dword v[2:3], v4, off
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX90A-GISEL-NEXT: global_store_dword v[2:3], v0, off
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: global_store_dword v[2:3], v0, off
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-LABEL: v_multi_use_mul_chain_add_other_use_all:
|
|
; GFX10: ; %bb.0: ; %bb
|
|
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v0, v2, v0
|
|
; GFX10-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v0, 1, v2
|
|
; GFX10-NEXT: v_mul_lo_u32 v5, v1, v0
|
|
; GFX10-NEXT: global_store_dword v[3:4], v2, off
|
|
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
|
|
; GFX10-NEXT: global_store_dword v[3:4], v1, off
|
|
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
|
|
; GFX10-NEXT: global_store_dword v[3:4], v5, off
|
|
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v0, v5, v0
|
|
; GFX10-NEXT: s_setpc_b64 s[30:31]
|
|
bb:
|
|
%i = add i32 %arg, 1
|
|
%i3 = mul i32 %i, %arg1
|
|
store volatile i32 %i3, ptr addrspace(1) %ptr
|
|
%i4 = add i32 %i3, %i
|
|
%i5 = mul i32 %i4, %arg1
|
|
store volatile i32 %i5, ptr addrspace(1) %ptr
|
|
%i6 = add i32 %i3, 1
|
|
%i7 = mul i32 %i5, %i6
|
|
store volatile i32 %i7, ptr addrspace(1) %ptr
|
|
%i8 = add i32 %i7, %i6
|
|
ret i32 %i8
|
|
}
|
|
|
|
define i32 @v_multi_use_mul_chain_add_other_use_some(i32 %arg, i32 %arg1, i32 %arg2, ptr addrspace(1) %ptr) {
|
|
; GFX6-SDAG-LABEL: v_multi_use_mul_chain_add_other_use_some:
|
|
; GFX6-SDAG: ; %bb.0: ; %bb
|
|
; GFX6-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX6-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX6-SDAG-NEXT: s_mov_b32 s7, 0xf000
|
|
; GFX6-SDAG-NEXT: s_mov_b32 s4, s6
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX6-SDAG-NEXT: s_mov_b32 s5, s6
|
|
; GFX6-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX6-SDAG-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
|
|
; GFX6-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX6-SDAG-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
|
|
; GFX6-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; GFX6-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v1
|
|
; GFX6-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX6-GISEL-LABEL: v_multi_use_mul_chain_add_other_use_some:
|
|
; GFX6-GISEL: ; %bb.0: ; %bb
|
|
; GFX6-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX6-GISEL-NEXT: s_mov_b32 s6, 0
|
|
; GFX6-GISEL-NEXT: s_mov_b32 s7, 0xf000
|
|
; GFX6-GISEL-NEXT: s_mov_b64 s[4:5], 0
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX6-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX6-GISEL-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
|
|
; GFX6-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX6-GISEL-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
|
|
; GFX6-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0)
|
|
; GFX6-GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v1
|
|
; GFX6-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX7-SDAG-LABEL: v_multi_use_mul_chain_add_other_use_some:
|
|
; GFX7-SDAG: ; %bb.0: ; %bb
|
|
; GFX7-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX7-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX7-SDAG-NEXT: s_mov_b32 s7, 0xf000
|
|
; GFX7-SDAG-NEXT: s_mov_b32 s4, s6
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX7-SDAG-NEXT: s_mov_b32 s5, s6
|
|
; GFX7-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX7-SDAG-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
|
|
; GFX7-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX7-SDAG-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
|
|
; GFX7-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX7-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v1
|
|
; GFX7-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX7-GISEL-LABEL: v_multi_use_mul_chain_add_other_use_some:
|
|
; GFX7-GISEL: ; %bb.0: ; %bb
|
|
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX7-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX7-GISEL-NEXT: s_mov_b32 s6, 0
|
|
; GFX7-GISEL-NEXT: s_mov_b32 s7, 0xf000
|
|
; GFX7-GISEL-NEXT: s_mov_b64 s[4:5], 0
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX7-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX7-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX7-GISEL-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
|
|
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX7-GISEL-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
|
|
; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX7-GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v1
|
|
; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-LABEL: v_multi_use_mul_chain_add_other_use_some:
|
|
; GFX8: ; %bb.0: ; %bb
|
|
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
; GFX8-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 1, v2
|
|
; GFX8-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-NEXT: flat_store_dword v[3:4], v2
|
|
; GFX8-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX8-NEXT: flat_store_dword v[3:4], v0
|
|
; GFX8-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
; GFX8-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-LABEL: v_multi_use_mul_chain_add_other_use_some:
|
|
; GFX900: ; %bb.0: ; %bb
|
|
; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX900-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX900-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX900-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-NEXT: global_store_dword v[3:4], v2, off
|
|
; GFX900-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX900-NEXT: global_store_dword v[3:4], v0, off
|
|
; GFX900-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX900-NEXT: v_add_u32_e32 v0, v0, v1
|
|
; GFX900-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: v_multi_use_mul_chain_add_other_use_some:
|
|
; GFX90A-SDAG: ; %bb.0: ; %bb
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX90A-SDAG-NEXT: v_mov_b32_e32 v5, v4
|
|
; GFX90A-SDAG-NEXT: v_mov_b32_e32 v4, v3
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-SDAG-NEXT: global_store_dword v[4:5], v2, off
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX90A-SDAG-NEXT: global_store_dword v[4:5], v0, off
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v0, v1
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: v_multi_use_mul_chain_add_other_use_some:
|
|
; GFX90A-GISEL: ; %bb.0: ; %bb
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_mov_b32_e32 v2, v3
|
|
; GFX90A-GISEL-NEXT: v_mov_b32_e32 v3, v4
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v4, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v4, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v4
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: global_store_dword v[2:3], v4, off
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX90A-GISEL-NEXT: global_store_dword v[2:3], v0, off
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-LABEL: v_multi_use_mul_chain_add_other_use_some:
|
|
; GFX10: ; %bb.0: ; %bb
|
|
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v0, v2, v0
|
|
; GFX10-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v1, 1, v2
|
|
; GFX10-NEXT: v_mul_lo_u32 v5, v0, v1
|
|
; GFX10-NEXT: global_store_dword v[3:4], v2, off
|
|
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
|
|
; GFX10-NEXT: global_store_dword v[3:4], v5, off
|
|
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v0, v5, v1
|
|
; GFX10-NEXT: s_setpc_b64 s[30:31]
|
|
bb:
|
|
%i = add i32 %arg, 1
|
|
%i3 = mul i32 %i, %arg1
|
|
store volatile i32 %i3, ptr addrspace(1) %ptr
|
|
%i4 = add i32 %i3, %i
|
|
%i5 = mul i32 %i4, %arg1
|
|
%i6 = add i32 %i3, 1
|
|
%i7 = mul i32 %i5, %i6
|
|
store volatile i32 %i7, ptr addrspace(1) %ptr
|
|
%i8 = add i32 %i7, %i6
|
|
ret i32 %i8
|
|
}
|
|
|
|
define i32 @clpeak_imad_pat_i32_x2(i32 %x, i32 %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_i32_x2:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v2, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v2, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v1
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_i32_x2:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v2, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v2, v1
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v2, v1
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_i32_x2:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, 1, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v2, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, 1, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v2, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v0, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v0, v1
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_i32_x2:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, v2, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, v2, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: clpeak_imad_pat_i32_x2:
|
|
; GFX900-SDAG: ; %bb.0: ; %entry
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, v2, v1
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, v2, v1
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v0, v2, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v1, v0, v[1:2]
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: clpeak_imad_pat_i32_x2:
|
|
; GFX900-GISEL: ; %bb.0: ; %entry
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, v2, v1
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, v2, v1
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: clpeak_imad_pat_i32_x2:
|
|
; GFX90A-SDAG: ; %bb.0: ; %entry
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, v2, v1
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, v2, v1
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v0, v2, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v0, v[2:3]
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: clpeak_imad_pat_i32_x2:
|
|
; GFX90A-GISEL: ; %bb.0: ; %entry
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v2, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v2, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v2, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v2
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_i32_x2:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, 1, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, v2, v1
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, 1, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, v2, v1
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[1:2], null, v0, v2, v[0:1]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v1, v0, v[1:2]
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_i32_x2:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, v2, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, v2, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v1, v0
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y38 = add i32 %x, 1
|
|
%add = mul i32 %y38, %y
|
|
%mul139 = add i32 %add, %y38
|
|
%add2 = mul i32 %mul139, %y
|
|
%add240 = add i32 %add, 1
|
|
%add4 = mul i32 %add2, %add240
|
|
%mul541 = add i32 %add4, %add240
|
|
%add6 = mul i32 %mul541, %add2
|
|
%add642 = add i32 %add4, 1
|
|
%add8 = mul i32 %add6, %add642
|
|
%mul943 = add i32 %add8, %add642
|
|
%add10 = mul i32 %mul943, %add6
|
|
%add1044 = add i32 %add8, 1
|
|
%add1246 = add i32 %add10, 1
|
|
%mul1345 = mul i32 %add10, %add1044
|
|
%add14 = mul i32 %mul1345, %add1246
|
|
ret i32 %add14
|
|
}
|
|
|
|
define <2 x i32> @clpeak_imad_pat_v2i32_x2(<2 x i32> %x, <2 x i32> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_v2i32_x2:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v5, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v4, v1, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v5, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v4, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 1, v5
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v3, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v4, vcc, 1, v4
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v5, v1, v4
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, v3, v2
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v4, vcc, v5, v4
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v4, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 1, v3
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v3, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v4, vcc, 1, v5
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v5, v1, v4
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, v3, v2
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, v5, v4
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v2, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v2, v0, v3
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v3, v1, v5
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, v2, v0
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, v3, v1
|
|
; GFX67-SDAG-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, v1, v3
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_v2i32_x2:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v4, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v5, v1
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v3, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v5, v1, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v3, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, v5, v4
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v4, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v3, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v5, v1, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v3, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, v5, v4
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v4, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v2, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX67-GISEL-NEXT: v_mul_lo_u32 v1, v2, v1
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_v2i32_x2:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, 1, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v5, v0, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v4, v1, v3
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v5, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v4, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, 1, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v3, v0, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v4, vcc, 1, v4
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v5, v1, v4
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v3, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v4, vcc, v5, v4
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v4, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, 1, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v3, v0, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v4, vcc, 1, v5
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v5, v1, v4
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v3, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v5, v4
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v2, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v2, v0, v3
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v3, v1, v5
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v2, vcc, v2, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v3, vcc, v3, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v0, vcc, v0, v2
|
|
; GFX8-SDAG-NEXT: v_add_u32_e32 v1, vcc, v1, v3
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_v2i32_x2:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v0, vcc, v4, v0
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, v5, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, 1, v4
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v4, vcc, 1, v5
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v3, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v5, v1, v4
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, v3, v2
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v4, vcc, v5, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v4, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, 1, v3
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v4, vcc, 1, v5
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v3, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v5, v1, v4
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, v3, v2
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v4, vcc, v5, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v4, v1
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v2, vcc, 1, v3
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v3, vcc, 1, v5
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v4, vcc, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v2, v1, v3
|
|
; GFX8-GISEL-NEXT: v_add_u32_e32 v1, vcc, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u32 v1, v2, v1
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: clpeak_imad_pat_v2i32_x2:
|
|
; GFX900-SDAG: ; %bb.0: ; %entry
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v0, v4, v0
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v1, v5, v1
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v3, v0, v2
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v4, 1, v5
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v5, v1, v4
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v2, v3, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v2, v5, v4
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v1, v2, v1
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v2, 1, v3
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v3, v0, v2
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v4, 1, v5
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v5, v1, v4
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v2, v3, v2
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX900-SDAG-NEXT: v_add_u32_e32 v2, v5, v4
|
|
; GFX900-SDAG-NEXT: v_mul_lo_u32 v2, v2, v1
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v0, v3, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v5, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v3, v0, v[3:4]
|
|
; GFX900-SDAG-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v4, v2, v[4:5]
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: clpeak_imad_pat_v2i32_x2:
|
|
; GFX900-GISEL: ; %bb.0: ; %entry
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, v4, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, v5, v1
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v3, 1, v5
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v2, v4, v2
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v3, v5, v3
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v3, 1, v5
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v2, v4, v2
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v3, v5, v3
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v3, 1, v5
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v2, v0, v2
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v3, v1, v3
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX900-GISEL-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX900-GISEL-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-SDAG-LABEL: clpeak_imad_pat_v2i32_x2:
|
|
; GFX90A-SDAG: ; %bb.0: ; %entry
|
|
; GFX90A-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v1, v5, v1
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v0, v4, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v3, 1, v5
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v4, v1, v3
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v5, v0, v2
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v2, v5, v2
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v3, v4, v3
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v3, 1, v5
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v6, v0, v3
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v4, v1, v2
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v2, v4, v2
|
|
; GFX90A-SDAG-NEXT: v_add_u32_e32 v3, v6, v3
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v0, v3, v0
|
|
; GFX90A-SDAG-NEXT: v_mul_lo_u32 v2, v2, v1
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v2, v4, v[2:3]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v0, v6, v[0:1]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v6, v0, v[6:7]
|
|
; GFX90A-SDAG-NEXT: v_mad_u64_u32 v[2:3], s[4:5], v4, v2, v[4:5]
|
|
; GFX90A-SDAG-NEXT: v_mov_b32_e32 v1, v2
|
|
; GFX90A-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX90A-GISEL-LABEL: clpeak_imad_pat_v2i32_x2:
|
|
; GFX90A-GISEL: ; %bb.0: ; %entry
|
|
; GFX90A-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, 1, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v0, v4, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v1, v5, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, 1, v5
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, v4, v2
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, v5, v3
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, 1, v5
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, v4, v2
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, v5, v3
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v2, 1, v4
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v3, 1, v5
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v4, 1, v0
|
|
; GFX90A-GISEL-NEXT: v_add_u32_e32 v5, 1, v1
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v0, v0, v4
|
|
; GFX90A-GISEL-NEXT: v_mul_lo_u32 v1, v1, v5
|
|
; GFX90A-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_v2i32_x2:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v0, v4, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v1, v5, v1
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v2, 1, v4
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v3, 1, v5
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v2, v4, v2
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v3, v5, v3
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v2, 1, v4
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v3, 1, v5
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v2, v4, v2
|
|
; GFX10-SDAG-NEXT: v_add_nc_u32_e32 v3, v5, v3
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u32 v2, v3, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[3:4], null, v0, v4, v[0:1]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[4:5], null, v2, v5, v[2:3]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[0:1], null, v3, v0, v[3:4]
|
|
; GFX10-SDAG-NEXT: v_mad_u64_u32 v[1:2], null, v4, v2, v[4:5]
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_v2i32_x2:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, v4, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, v5, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v1, v3
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, 1, v4
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v3, 1, v5
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, v4, v2
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v3, v5, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, 1, v4
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v3, 1, v5
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v4, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v5, v1, v3
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, v4, v2
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v3, v5, v3
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v2, 1, v4
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v3, 1, v5
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v2, v0, v2
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v3, v1, v3
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v0, 1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u32_e32 v1, 1, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u32 v1, v3, v1
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y38 = add <2 x i32> %x, <i32 1, i32 1>
|
|
%add = mul <2 x i32> %y38, %y
|
|
%mul139 = add <2 x i32> %add, %y38
|
|
%add2 = mul <2 x i32> %mul139, %y
|
|
%add240 = add <2 x i32> %add, <i32 1, i32 1>
|
|
%add4 = mul <2 x i32> %add2, %add240
|
|
%mul541 = add <2 x i32> %add4, %add240
|
|
%add6 = mul <2 x i32> %mul541, %add2
|
|
%add642 = add <2 x i32> %add4, <i32 1, i32 1>
|
|
%add8 = mul <2 x i32> %add6, %add642
|
|
%mul943 = add <2 x i32> %add8, %add642
|
|
%add10 = mul <2 x i32> %mul943, %add6
|
|
%add1044 = add <2 x i32> %add8, <i32 1, i32 1>
|
|
%add1246 = add <2 x i32> %add10, <i32 1, i32 1>
|
|
%mul1345 = mul <2 x i32> %add10, %add1044
|
|
%add14 = mul <2 x i32> %mul1345, %add1246
|
|
ret <2 x i32> %add14
|
|
}
|
|
|
|
define signext i16 @clpeak_imad_pat_i16_x2(i16 signext %x, i16 signext %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_i16_x2:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v1, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v1, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v1, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_i16_x2:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_i16_x2:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_i16_x2:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v3, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v3, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v0, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v3, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v3, v2, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v4, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v0, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v2, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v1, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v1, v4, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v1
|
|
; GFX8-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_imad_pat_i16_x2:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_imad_pat_i16_x2:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v3, 1, v1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v4, v3, v2
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v1, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v1, v0, v4
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v2, v3, v2, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v3, v2, v1
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v4, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v4, v0, v3
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v1, v2, v1, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v2, v1, v4
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v3, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v1, v1, v4, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_i16_x2:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_i16_x2:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v2, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v2, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%conv69 = add i16 %x, 1
|
|
%add = mul i16 %conv69, %y
|
|
%conv470 = add i16 %y, 1
|
|
%add8 = mul i16 %conv470, %add
|
|
%conv1071 = add i16 %add, 1
|
|
%add14 = mul i16 %conv1071, %add8
|
|
%conv1672 = add i16 %add8, 1
|
|
%add20 = mul i16 %conv1672, %add14
|
|
%conv2273 = add i16 %add14, 1
|
|
%add26 = mul i16 %conv2273, %add20
|
|
%conv2874 = add i16 %add20, 1
|
|
%add32 = mul i16 %conv2874, %add26
|
|
%conv3475 = add i16 %add26, 1
|
|
%add38 = mul i16 %conv3475, %add32
|
|
%conv4076 = add i16 %add32, 1
|
|
%add44 = mul i16 %add38, %conv4076
|
|
ret i16 %add44
|
|
}
|
|
|
|
define zeroext i16 @clpeak_umad_pat_i16_x2(i16 zeroext %x, i16 zeroext %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_umad_pat_i16_x2:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v2, v0, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v3, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v3, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v1, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v0, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v3, v2, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v2, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v1, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v0, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v2, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v2, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v1, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v2, v3, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_umad_pat_i16_x2:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_umad_pat_i16_x2:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v1, v0
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_umad_pat_i16_x2:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v0, v1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v3, 1, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v3, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v0, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v3, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v3, v2, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v4, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v0, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v2, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v1, v4
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v1, v4, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v1
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_umad_pat_i16_x2:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v1, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v1, v0, v1
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v0, v1, v0
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_umad_pat_i16_x2:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v0, 1, v0
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_add_u16_e32 v3, 1, v1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v4, v3, v2
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v1, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v1, v0, v4
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v2, v3, v2, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v3, v2, v1
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v4, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v4, v0, v3
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v1, v2, v1, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v2, v1, v4
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v3, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v1, v1, v4, 1
|
|
; GFX9-GISEL-NEXT: v_mul_lo_u16_e32 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_umad_pat_i16_x2:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v1, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v1, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_umad_pat_i16_x2:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v2, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v2, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v1, v0
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v1, 1
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%conv69 = add i16 %x, 1
|
|
%add = mul i16 %conv69, %y
|
|
%conv470 = add i16 %y, 1
|
|
%add8 = mul i16 %conv470, %add
|
|
%conv1071 = add i16 %add, 1
|
|
%add14 = mul i16 %conv1071, %add8
|
|
%conv1672 = add i16 %add8, 1
|
|
%add20 = mul i16 %conv1672, %add14
|
|
%conv2273 = add i16 %add14, 1
|
|
%add26 = mul i16 %conv2273, %add20
|
|
%conv2874 = add i16 %add20, 1
|
|
%add32 = mul i16 %conv2874, %add26
|
|
%conv3475 = add i16 %add26, 1
|
|
%add38 = mul i16 %conv3475, %add32
|
|
%conv4076 = add i16 %add32, 1
|
|
%add44 = mul i16 %add38, %conv4076
|
|
ret i16 %add44
|
|
}
|
|
|
|
define <2 x i16> @clpeak_imad_pat_v2i16_x2(<2 x i16> %x, <2 x i16> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_imad_pat_v2i16_x2:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v4, v3, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v5, v2, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v4, v3, 1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v5, v2, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v1, v4, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v0, v5, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v3, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v1, v4, 1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v2, v2, v0
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v5, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v3, v4, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v6, v3, v4
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v2, v5, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v2, v5, 1
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v5, 16, v6
|
|
; GFX67-SDAG-NEXT: s_mov_b32 s4, 0x10000
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v5, vcc, s4, v5
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, 1
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v2, 16, v1
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v5, 16, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 0x10000, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v4, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v5
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v2, 16, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v3, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_imad_pat_v2i16_x2:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v5, v5, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v4, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v5, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v4, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v5, v1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v0, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, v5, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v4, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v3, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v2, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v5, v1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v0, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, v5, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v4, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v3, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v2, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v4, 16, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v2, 16, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_imad_pat_v2i16_x2:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mov_b32_e32 v3, 1
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v2, 1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v3, 16, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v0, v3, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v5, v2, v1, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v4, v4, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v3, 1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v5, v5, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v2, v1, 1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v4, v0, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v3, v5, v1, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v2, v2, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v4, v0, 1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v3, v3, v5
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v5, v1, 1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v4, v2, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v2, v0, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v5, v3, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v3, v1, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v0, v0, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v1, v1, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v0, v4, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v3, v1, v5, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v2, v0, v2
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v3, v1, v3
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_imad_pat_v2i16_x2:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v3, 1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v2, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v2, v1, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v0, v3, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v4, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v5, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v2, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v4, v1, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v5, v0, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v2, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v3, v3, v5
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v4, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v5, v0, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v2, v1, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v3, v0, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v6, v4, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v7, v5, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v2, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v3, v0, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v4, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v5, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v6, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v7, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v1, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_imad_pat_v2i16_x2:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v2, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_imad_pat_v2i16_x2:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v2, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_imad_pat_v2i16_x2:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v2, v2, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v2, v1, v2
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v1, v1, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v2, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_imad_pat_v2i16_x2:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v2, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v1, v2
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v1, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y38 = add <2 x i16> %x, <i16 1, i16 1>
|
|
%add = mul <2 x i16> %y38, %y
|
|
%mul139 = add <2 x i16> %add, %y38
|
|
%add2 = mul <2 x i16> %mul139, %y
|
|
%add240 = add <2 x i16> %add, <i16 1, i16 1>
|
|
%add4 = mul <2 x i16> %add2, %add240
|
|
%mul541 = add <2 x i16> %add4, %add240
|
|
%add6 = mul <2 x i16> %mul541, %add2
|
|
%add642 = add <2 x i16> %add4, <i16 1, i16 1>
|
|
%add8 = mul <2 x i16> %add6, %add642
|
|
%mul943 = add <2 x i16> %add8, %add642
|
|
%add10 = mul <2 x i16> %mul943, %add6
|
|
%add1044 = add <2 x i16> %add8, <i16 1, i16 1>
|
|
%add1246 = add <2 x i16> %add10, <i16 1, i16 1>
|
|
%mul1345 = mul <2 x i16> %add10, %add1044
|
|
%add14 = mul <2 x i16> %mul1345, %add1246
|
|
ret <2 x i16> %add14
|
|
}
|
|
|
|
define <2 x i16> @clpeak_umad_pat_v2i16_x2(<2 x i16> %x, <2 x i16> %y) {
|
|
; GFX67-SDAG-LABEL: clpeak_umad_pat_v2i16_x2:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v4, v3, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v5, v2, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v4, v3, 1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v5, v2, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v1, v4, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v0, v5, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v3, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v1, v4, 1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v2, v2, v0
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v5, 1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v5, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v3, v4, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v6, v3, v4
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v2, v5, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v2, v5, 1
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v5, 16, v6
|
|
; GFX67-SDAG-NEXT: s_mov_b32 s4, 0x10000
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v0, v2
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v5, vcc, s4, v5
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, 1
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v2, 16, v1
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v5, 16, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_add_i32_e32 v2, vcc, 0x10000, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v3, v4, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v5
|
|
; GFX67-SDAG-NEXT: v_lshrrev_b32_e32 v2, 16, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v0, v3, v0
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: clpeak_umad_pat_v2i16_x2:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v5, v5, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v4, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v5, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v4, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v5, v1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v0, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, v5, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v4, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v3, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v2, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v5, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v5, 16, v5
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v4, v4, v5
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v5, v1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v4, v0, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, v5, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v4, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v3, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v5
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v2, v0
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v4, vcc, 1, v1
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, 1, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v3, v3, v4
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v4, 16, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v4
|
|
; GFX67-GISEL-NEXT: v_lshrrev_b32_e32 v2, 16, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: clpeak_umad_pat_v2i16_x2:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mov_b32_e32 v3, 1
|
|
; GFX8-SDAG-NEXT: v_add_u16_e32 v2, 1, v0
|
|
; GFX8-SDAG-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v3, 16, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v0, v3, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v5, v2, v1, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v4, v4, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v3, 1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v5, v5, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v2, v1, 1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v4, v0, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v3, v5, v1, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v2, v2, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v4, v0, 1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v3, v3, v5
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v5, v1, 1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v4, v2, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v2, v0, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v5, v3, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v3, v1, v1
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v0, v0, v2
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v1, v1, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v0, v4, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v3, v1, v5, v1
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v2, v0, v2
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v1, v3, v1, v3
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: clpeak_umad_pat_v2i16_x2:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v3, 1
|
|
; GFX8-GISEL-NEXT: v_add_u16_e32 v2, 1, v0
|
|
; GFX8-GISEL-NEXT: v_add_u16_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v2, v1, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v0, v3, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v4, v4, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v5, v5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v2, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v4, v1, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v5, v0, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v2, v2, v4
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v3, v3, v5
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v4, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v5, v0, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v4, v2, v1, v1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v5, v3, v0, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v6, v4, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v7, v5, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v2, v1, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v3, v0, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v4, v2, 1
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v3, v5, v3, 1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v6, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v0, v7, v0
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v1, v1, v2
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v1, v0
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: clpeak_umad_pat_v2i16_x2:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_sub_u16 v2, v0, -1 op_sel_hi:[1,0]
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: clpeak_umad_pat_v2i16_x2:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_add_u16 v2, v0, 1 op_sel_hi:[1,0]
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX9-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: clpeak_umad_pat_v2i16_x2:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v0, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v2, v2, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v1, v0, v2
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v2, v1, v2
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v1, v1, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v2, v0
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v1, v2, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_sub_u16 v2, v0, -1 op_sel_hi:[1,0]
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-SDAG-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: clpeak_umad_pat_v2i16_x2:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v2, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v1, v0, v2
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v1, v2
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v1, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v2, v0
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v2, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v2, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v1, v0
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v1, v2, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_add_u16 v2, v0, 1 op_sel_hi:[1,0]
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_pk_mul_lo_u16 v0, v0, v2
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%y38 = add <2 x i16> %x, <i16 1, i16 1>
|
|
%add = mul <2 x i16> %y38, %y
|
|
%mul139 = add <2 x i16> %add, %y38
|
|
%add2 = mul <2 x i16> %mul139, %y
|
|
%add240 = add <2 x i16> %add, <i16 1, i16 1>
|
|
%add4 = mul <2 x i16> %add2, %add240
|
|
%mul541 = add <2 x i16> %add4, %add240
|
|
%add6 = mul <2 x i16> %mul541, %add2
|
|
%add642 = add <2 x i16> %add4, <i16 1, i16 1>
|
|
%add8 = mul <2 x i16> %add6, %add642
|
|
%mul943 = add <2 x i16> %add8, %add642
|
|
%add10 = mul <2 x i16> %mul943, %add6
|
|
%add1044 = add <2 x i16> %add8, <i16 1, i16 1>
|
|
%add1246 = add <2 x i16> %add10, <i16 1, i16 1>
|
|
%mul1345 = mul <2 x i16> %add10, %add1044
|
|
%add14 = mul <2 x i16> %mul1345, %add1246
|
|
ret <2 x i16> %add14
|
|
}
|
|
|
|
; Multiple uses of mul with different variable addend
|
|
define <2 x i32> @multi_use_mul_mad_i32_var(i32 %x, i32 %y, i32 %z0, i32 %z1) {
|
|
; GFX67-LABEL: multi_use_mul_mad_i32_var:
|
|
; GFX67: ; %bb.0: ; %entry
|
|
; GFX67-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX67-NEXT: v_add_i32_e32 v0, vcc, v1, v2
|
|
; GFX67-NEXT: v_add_i32_e32 v1, vcc, v1, v3
|
|
; GFX67-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-LABEL: multi_use_mul_mad_i32_var:
|
|
; GFX8: ; %bb.0: ; %entry
|
|
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v1, v2
|
|
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v3
|
|
; GFX8-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: multi_use_mul_mad_i32_var:
|
|
; GFX9: ; %bb.0: ; %entry
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX9-NEXT: v_add_u32_e32 v0, v1, v2
|
|
; GFX9-NEXT: v_add_u32_e32 v1, v1, v3
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-LABEL: multi_use_mul_mad_i32_var:
|
|
; GFX10: ; %bb.0: ; %entry
|
|
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v0, v1, v2
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v1, v1, v3
|
|
; GFX10-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%mul = mul i32 %x, %y
|
|
%add0 = add i32 %mul, %z0
|
|
%add1 = add i32 %mul, %z1
|
|
%insert.0 = insertelement <2 x i32> poison, i32 %add0, i32 0
|
|
%insert.1 = insertelement <2 x i32> %insert.0, i32 %add1, i32 1
|
|
ret <2 x i32> %insert.1
|
|
}
|
|
|
|
define <2 x i16> @multi_use_mul_mad_i16_var(i16 %x, i16 %y, i16 %z0, i16 %z1) {
|
|
; GFX67-SDAG-LABEL: multi_use_mul_mad_i16_var:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v4, v1, v2
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v4, v1, v3
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: multi_use_mul_mad_i16_var:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v1, v2
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v3
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: multi_use_mul_mad_i16_var:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v3, v0, v1, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v1, v2
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v3
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: multi_use_mul_mad_i16_var:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v0, v1, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v1, v3
|
|
; GFX8-GISEL-NEXT: v_lshlrev_b32_e32 v0, 16, v0
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v2, v0
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-SDAG-LABEL: multi_use_mul_mad_i16_var:
|
|
; GFX9-SDAG: ; %bb.0: ; %entry
|
|
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v2, v0, v1, v2
|
|
; GFX9-SDAG-NEXT: v_mad_legacy_u16 v0, v0, v1, v3
|
|
; GFX9-SDAG-NEXT: s_mov_b32 s4, 0x5040100
|
|
; GFX9-SDAG-NEXT: v_perm_b32 v0, v0, v2, s4
|
|
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-GISEL-LABEL: multi_use_mul_mad_i16_var:
|
|
; GFX9-GISEL: ; %bb.0: ; %entry
|
|
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v2, v0, v1, v2
|
|
; GFX9-GISEL-NEXT: v_mad_legacy_u16 v0, v0, v1, v3
|
|
; GFX9-GISEL-NEXT: v_lshl_or_b32 v0, v0, 16, v2
|
|
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: multi_use_mul_mad_i16_var:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v2, v0, v1, v2
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v0, v1, v3
|
|
; GFX10-SDAG-NEXT: v_perm_b32 v0, v0, v2, 0x5040100
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: multi_use_mul_mad_i16_var:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v0, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v1, v0, v2
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v0, v3
|
|
; GFX10-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX10-GISEL-NEXT: v_lshl_or_b32 v0, v0, 16, v1
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%mul = mul i16 %x, %y
|
|
%add0 = add i16 %mul, %z0
|
|
%add1 = add i16 %mul, %z1
|
|
%insert.0 = insertelement <2 x i16> poison, i16 %add0, i16 0
|
|
%insert.1 = insertelement <2 x i16> %insert.0, i16 %add1, i16 1
|
|
ret <2 x i16> %insert.1
|
|
}
|
|
|
|
define i32 @other_use_mul_mad_i32_var(i32 %x, i32 %y, i32 %z, ptr addrspace(3) %ptr) {
|
|
; GFX67-LABEL: other_use_mul_mad_i32_var:
|
|
; GFX67: ; %bb.0: ; %entry
|
|
; GFX67-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX67-NEXT: s_mov_b32 m0, -1
|
|
; GFX67-NEXT: v_add_i32_e32 v0, vcc, v1, v2
|
|
; GFX67-NEXT: ds_write_b32 v3, v1
|
|
; GFX67-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX67-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-LABEL: other_use_mul_mad_i32_var:
|
|
; GFX8: ; %bb.0: ; %entry
|
|
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX8-NEXT: s_mov_b32 m0, -1
|
|
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v1, v2
|
|
; GFX8-NEXT: ds_write_b32 v3, v1
|
|
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX8-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: other_use_mul_mad_i32_var:
|
|
; GFX9: ; %bb.0: ; %entry
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX9-NEXT: v_add_u32_e32 v0, v1, v2
|
|
; GFX9-NEXT: ds_write_b32 v3, v1
|
|
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-LABEL: other_use_mul_mad_i32_var:
|
|
; GFX10: ; %bb.0: ; %entry
|
|
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-NEXT: v_mul_lo_u32 v1, v0, v1
|
|
; GFX10-NEXT: v_add_nc_u32_e32 v0, v1, v2
|
|
; GFX10-NEXT: ds_write_b32 v3, v1
|
|
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX10-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%mul = mul i32 %x, %y
|
|
%add0 = add i32 %mul, %z
|
|
store i32 %mul, ptr addrspace(3) %ptr
|
|
ret i32 %add0
|
|
}
|
|
|
|
define i16 @other_use_mul_mad_i16_var(i16 %x, i16 %y, i16 %z, ptr addrspace(3) %ptr) {
|
|
; GFX67-SDAG-LABEL: other_use_mul_mad_i16_var:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v4, v0, v1
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v1, v2
|
|
; GFX67-SDAG-NEXT: s_mov_b32 m0, -1
|
|
; GFX67-SDAG-NEXT: ds_write_b16 v3, v4
|
|
; GFX67-SDAG-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: other_use_mul_mad_i16_var:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v1, v2
|
|
; GFX67-GISEL-NEXT: s_mov_b32 m0, -1
|
|
; GFX67-GISEL-NEXT: ds_write_b16 v3, v1
|
|
; GFX67-GISEL-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-LABEL: other_use_mul_mad_i16_var:
|
|
; GFX8: ; %bb.0: ; %entry
|
|
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-NEXT: v_mul_lo_u16_e32 v4, v0, v1
|
|
; GFX8-NEXT: v_mad_u16 v0, v0, v1, v2
|
|
; GFX8-NEXT: s_mov_b32 m0, -1
|
|
; GFX8-NEXT: ds_write_b16 v3, v4
|
|
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX8-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: other_use_mul_mad_i16_var:
|
|
; GFX9: ; %bb.0: ; %entry
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_mul_lo_u16_e32 v4, v0, v1
|
|
; GFX9-NEXT: v_mad_legacy_u16 v0, v0, v1, v2
|
|
; GFX9-NEXT: ds_write_b16 v3, v4
|
|
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-SDAG-LABEL: other_use_mul_mad_i16_var:
|
|
; GFX10-SDAG: ; %bb.0: ; %entry
|
|
; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: v_mul_lo_u16 v4, v0, v1
|
|
; GFX10-SDAG-NEXT: v_mad_u16 v0, v0, v1, v2
|
|
; GFX10-SDAG-NEXT: ds_write_b16 v3, v4
|
|
; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-GISEL-LABEL: other_use_mul_mad_i16_var:
|
|
; GFX10-GISEL: ; %bb.0: ; %entry
|
|
; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: v_mul_lo_u16 v1, v0, v1
|
|
; GFX10-GISEL-NEXT: v_add_nc_u16 v0, v1, v2
|
|
; GFX10-GISEL-NEXT: ds_write_b16 v3, v1
|
|
; GFX10-GISEL-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%mul = mul i16 %x, %y
|
|
%add0 = add i16 %mul, %z
|
|
store i16 %mul, ptr addrspace(3) %ptr
|
|
ret i16 %add0
|
|
}
|
|
|
|
define <4 x i16> @multi_use_mul_mad_v2i16_var(<2 x i16> %x, <2 x i16> %y, <2 x i16> %z0, <2 x i16> %z1) {
|
|
; GFX67-SDAG-LABEL: multi_use_mul_mad_v2i16_var:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v8, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v5, v1, v3, v5
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v4, v8, v2, v4
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v3, v1, v3, v7
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v2, v8, v2, v6
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v0, 16, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v4, 0xffff, v4
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v3
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v0, v4, v0
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v2, v2, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v5
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: multi_use_mul_mad_v2i16_var:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v2, v0, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v3, v0, v1
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v2, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v3, v5
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v2, vcc, v2, v6
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v7
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: multi_use_mul_mad_v2i16_var:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v4, 16, v1
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v5, 16, v0
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v6, 16, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v6, v5, v4, v6
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v2, v0, v1, v2
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v2, v2, v6
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v6, 16, v3
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v5, v4, v6
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v1, v3
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v1, v0, v4
|
|
; GFX8-SDAG-NEXT: v_mov_b32_e32 v0, v2
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: multi_use_mul_mad_v2i16_var:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v4, 16, v0
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v1
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v6, 16, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v6, v4, v5, v6
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v2, v0, v1, v2
|
|
; GFX8-GISEL-NEXT: v_lshlrev_b32_e32 v6, 16, v6
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v2, v2, v6
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v6, 16, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v1, v3
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v4, v5, v6
|
|
; GFX8-GISEL-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v1, v0, v1
|
|
; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, v2
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: multi_use_mul_mad_v2i16_var:
|
|
; GFX9: ; %bb.0: ; %entry
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_pk_mul_lo_u16 v1, v0, v1
|
|
; GFX9-NEXT: v_pk_add_u16 v0, v1, v2
|
|
; GFX9-NEXT: v_pk_add_u16 v1, v1, v3
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-LABEL: multi_use_mul_mad_v2i16_var:
|
|
; GFX10: ; %bb.0: ; %entry
|
|
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-NEXT: v_pk_mul_lo_u16 v1, v0, v1
|
|
; GFX10-NEXT: v_pk_add_u16 v0, v1, v2
|
|
; GFX10-NEXT: v_pk_add_u16 v1, v1, v3
|
|
; GFX10-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%mul = mul <2 x i16> %x, %y
|
|
%add0 = add <2 x i16> %mul, %z0
|
|
%add1 = add <2 x i16> %mul, %z1
|
|
%shuffle = shufflevector <2 x i16> %add0, <2 x i16> %add1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
|
|
ret <4 x i16> %shuffle
|
|
}
|
|
|
|
define <2 x i16> @other_use_mul_mad_v2i16_var(<2 x i16> %x, <2 x i16> %y, <2 x i16> %z, ptr addrspace(3) %ptr) {
|
|
; GFX67-SDAG-LABEL: other_use_mul_mad_v2i16_var:
|
|
; GFX67-SDAG: ; %bb.0: ; %entry
|
|
; GFX67-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v3, 0xffff, v3
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v7, v0, v2
|
|
; GFX67-SDAG-NEXT: v_mul_u32_u24_e32 v8, v1, v3
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v1, v1, v3, v5
|
|
; GFX67-SDAG-NEXT: v_mad_u32_u24 v0, v0, v2, v4
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v8, 16, v8
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v7, 0xffff, v7
|
|
; GFX67-SDAG-NEXT: v_lshlrev_b32_e32 v3, 16, v1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v7, v7, v8
|
|
; GFX67-SDAG-NEXT: v_or_b32_e32 v0, v0, v3
|
|
; GFX67-SDAG-NEXT: s_mov_b32 m0, -1
|
|
; GFX67-SDAG-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-SDAG-NEXT: ds_write_b32 v6, v7
|
|
; GFX67-SDAG-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX67-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX67-GISEL-LABEL: other_use_mul_mad_v2i16_var:
|
|
; GFX67-GISEL: ; %bb.0: ; %entry
|
|
; GFX67-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v2
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v0, v0, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v1, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v3
|
|
; GFX67-GISEL-NEXT: v_mul_u32_u24_e32 v1, v1, v2
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v3, 0xffff, v1
|
|
; GFX67-GISEL-NEXT: v_and_b32_e32 v2, 0xffff, v0
|
|
; GFX67-GISEL-NEXT: v_lshlrev_b32_e32 v3, 16, v3
|
|
; GFX67-GISEL-NEXT: v_or_b32_e32 v2, v2, v3
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v4
|
|
; GFX67-GISEL-NEXT: v_add_i32_e32 v1, vcc, v1, v5
|
|
; GFX67-GISEL-NEXT: s_mov_b32 m0, -1
|
|
; GFX67-GISEL-NEXT: ds_write_b32 v6, v2
|
|
; GFX67-GISEL-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX67-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-SDAG-LABEL: other_use_mul_mad_v2i16_var:
|
|
; GFX8-SDAG: ; %bb.0: ; %entry
|
|
; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v4, 16, v1
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v5, 16, v0
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_sdwa v6, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-SDAG-NEXT: v_mul_lo_u16_e32 v7, v0, v1
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v6, v7, v6
|
|
; GFX8-SDAG-NEXT: v_lshrrev_b32_e32 v7, 16, v2
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v4, v5, v4, v7
|
|
; GFX8-SDAG-NEXT: v_lshlrev_b32_e32 v4, 16, v4
|
|
; GFX8-SDAG-NEXT: v_mad_u16 v0, v0, v1, v2
|
|
; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v0, v4
|
|
; GFX8-SDAG-NEXT: s_mov_b32 m0, -1
|
|
; GFX8-SDAG-NEXT: ds_write_b32 v3, v6
|
|
; GFX8-SDAG-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX8-GISEL-LABEL: other_use_mul_mad_v2i16_var:
|
|
; GFX8-GISEL: ; %bb.0: ; %entry
|
|
; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v4, 16, v0
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v5, 16, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_e32 v6, v0, v1
|
|
; GFX8-GISEL-NEXT: v_mul_lo_u16_sdwa v7, v4, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v6, v6, v7
|
|
; GFX8-GISEL-NEXT: v_lshrrev_b32_e32 v7, 16, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v0, v0, v1, v2
|
|
; GFX8-GISEL-NEXT: v_mad_u16 v1, v4, v5, v7
|
|
; GFX8-GISEL-NEXT: v_lshlrev_b32_e32 v1, 16, v1
|
|
; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v0, v1
|
|
; GFX8-GISEL-NEXT: s_mov_b32 m0, -1
|
|
; GFX8-GISEL-NEXT: ds_write_b32 v3, v6
|
|
; GFX8-GISEL-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX9-LABEL: other_use_mul_mad_v2i16_var:
|
|
; GFX9: ; %bb.0: ; %entry
|
|
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX9-NEXT: v_pk_mul_lo_u16 v1, v0, v1
|
|
; GFX9-NEXT: v_pk_add_u16 v0, v1, v2
|
|
; GFX9-NEXT: ds_write_b32 v3, v1
|
|
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX9-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX10-LABEL: other_use_mul_mad_v2i16_var:
|
|
; GFX10: ; %bb.0: ; %entry
|
|
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX10-NEXT: v_pk_mul_lo_u16 v1, v0, v1
|
|
; GFX10-NEXT: v_pk_add_u16 v0, v1, v2
|
|
; GFX10-NEXT: ds_write_b32 v3, v1
|
|
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
|
|
; GFX10-NEXT: s_setpc_b64 s[30:31]
|
|
entry:
|
|
%mul = mul <2 x i16> %x, %y
|
|
%add0 = add <2 x i16> %mul, %z
|
|
store <2 x i16> %mul, ptr addrspace(3) %ptr
|
|
ret <2 x i16> %add0
|
|
}
|
|
|
|
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
|
|
; GFX6: {{.*}}
|
|
; GFX7: {{.*}}
|
|
; GFX90A: {{.*}}
|