# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -run-pass=si-fold-operands -verify-machineinstrs -o - %s | FileCheck --check-prefix=GFX90A %s
# Should fold only the aligned agpr copies with their uses.
---
name: aligned_agpr_64
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: aligned_agpr_64
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_64_align2 = IMPLICIT_DEF
; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]], 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_64_align2 = IMPLICIT_DEF
%2:areg_64_align2 = COPY killed %1
GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
...
---
name: unaligned_agpr_64
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: unaligned_agpr_64
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_64 = IMPLICIT_DEF
; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_64_align2 = COPY killed [[DEF]]
; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_64 = IMPLICIT_DEF
%2:areg_64_align2 = COPY killed %1
GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
...
---
name: aligned_agpr_96_sub0_subg1
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: aligned_agpr_96_sub0_subg1
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96_align2 = IMPLICIT_DEF
; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub0_sub1, 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_96_align2 = IMPLICIT_DEF
%2:areg_64_align2 = COPY killed %1.sub0_sub1:areg_96_align2
GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
...
---
name: unaligned_agpr_96_sub1_sub2
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: unaligned_agpr_96_sub1_sub2
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96_align2 = IMPLICIT_DEF
; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_64_align2 = COPY killed [[DEF]].sub1_sub2
; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_96_align2 = IMPLICIT_DEF
%2:areg_64_align2 = COPY killed %1.sub1_sub2:areg_96_align2
GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
...
---
name: aligned_agpr_96
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: aligned_agpr_96
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96_align2 = IMPLICIT_DEF
; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]], 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_96_align2 = IMPLICIT_DEF
%2:areg_96_align2 = COPY killed %1
GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
...
---
name: unaligned_agpr_96
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: unaligned_agpr_96
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96 = IMPLICIT_DEF
; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_96_align2 = COPY killed [[DEF]]
; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[COPY1]], 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_96 = IMPLICIT_DEF
%2:areg_96_align2 = COPY killed %1
GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
...
---
name: aligned_agpr_128_sub0_sub1
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: aligned_agpr_128_sub0_sub1
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub0_sub1, 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_128_align2 = IMPLICIT_DEF
%2:areg_64_align2 = COPY killed %1.sub0_sub1:areg_128_align2
GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
...
---
name: aligned_agpr_128_sub2_sub3
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: aligned_agpr_128_sub2_sub3
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub2_sub3, 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_128_align2 = IMPLICIT_DEF
%2:areg_64_align2 = COPY killed %1.sub2_sub3:areg_128_align2
GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
...
---
name: unaligned_agpr_128_sub1_sub2
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: unaligned_agpr_128_sub1_sub2
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_64_align2 = COPY killed [[DEF]].sub1_sub2
; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_128_align2 = IMPLICIT_DEF
%2:areg_64_align2 = COPY killed %1.sub1_sub2:areg_128_align2
GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
...
---
name: aligned_agpr_128_sub0_sub1_sub2
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: aligned_agpr_128_sub0_sub1_sub2
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]].sub0_sub1_sub2, 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_128_align2 = IMPLICIT_DEF
%2:areg_96_align2 = COPY killed %1.sub0_sub1_sub2:areg_128_align2
GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
...
---
name: unaligned_agpr_128_sub1_sub2_sub3
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: unaligned_agpr_128_sub1_sub2_sub3
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_96_align2 = COPY killed [[DEF]].sub1_sub2_sub3
; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[COPY1]], 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_128_align2 = IMPLICIT_DEF
%2:areg_96_align2 = COPY killed %1.sub1_sub2_sub3:areg_128_align2
GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
...
---
name: aligned_agpr_128
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: aligned_agpr_128
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
; GFX90A-NEXT: GLOBAL_STORE_DWORDX4 [[COPY]], [[DEF]], 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_128_align2 = IMPLICIT_DEF
%2:areg_128_align2 = COPY killed %1
GLOBAL_STORE_DWORDX4 %0, %2, 0, 0, implicit $exec
...
---
name: unaligned_agpr_128
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX90A-LABEL: name: unaligned_agpr_128
; GFX90A: liveins: $vgpr0_vgpr1
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128 = IMPLICIT_DEF
; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_128_align2 = COPY killed [[DEF]]
; GFX90A-NEXT: GLOBAL_STORE_DWORDX4 [[COPY]], [[COPY1]], 0, 0, implicit $exec
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:areg_128 = IMPLICIT_DEF
%2:areg_128_align2 = COPY killed %1
GLOBAL_STORE_DWORDX4 %0, %2, 0, 0, implicit $exec
...
Codebase Browser
llvm