# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 2
# RUN: llc -mtriple=amdgcn -mcpu=gfx1013 -verify-machineinstrs -run-pass si-load-store-opt -o - %s | FileCheck %s -check-prefixes=CHECK,GFX10
# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs -run-pass si-load-store-opt -o - %s | FileCheck %s -check-prefixes=CHECK,GFX11
# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs -run-pass si-load-store-opt -o - %s | FileCheck %s -check-prefixes=CHECK,GFX12
---
name: merge_s_load_x1_x1
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x1_x1
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %3:sreg_64_xexec = S_LOAD_DWORDX2_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s64), align 4)
; GFX10-NEXT: [[COPY:%[0-9]+]]:sreg_32_xm0_xexec = COPY %3.sub0
; GFX10-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed %3.sub1
;
; GFX11-LABEL: name: merge_s_load_x1_x1
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX2_IMM:%[0-9]+]]:sreg_64_xexec = S_LOAD_DWORDX2_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s64), align 4)
; GFX11-NEXT: [[COPY:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[S_LOAD_DWORDX2_IMM]].sub0
; GFX11-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_LOAD_DWORDX2_IMM]].sub1
;
; GFX12-LABEL: name: merge_s_load_x1_x1
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX2_IMM:%[0-9]+]]:sreg_64_xexec = S_LOAD_DWORDX2_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s64), align 4)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[S_LOAD_DWORDX2_IMM]].sub0
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_LOAD_DWORDX2_IMM]].sub1
%0:sgpr_64 = IMPLICIT_DEF
%1:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s32))
%2:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 4, 0 :: (dereferenceable invariant load (s32))
...
---
name: merge_s_load_x1_x2
body: |
bb.0:
; CHECK-LABEL: name: merge_s_load_x1_x2
; CHECK: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; CHECK-NEXT: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s32))
; CHECK-NEXT: [[S_LOAD_DWORDX2_IMM:%[0-9]+]]:sgpr_64 = S_LOAD_DWORDX2_IMM [[DEF]], 4, 0 :: (dereferenceable invariant load (s64))
%0:sgpr_64 = IMPLICIT_DEF
%1:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s32))
%2:sgpr_64 = S_LOAD_DWORDX2_IMM %0:sgpr_64, 4, 0 :: (dereferenceable invariant load (s64))
...
---
name: merge_s_load_x1_x3
body: |
bb.0:
; CHECK-LABEL: name: merge_s_load_x1_x3
; CHECK: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; CHECK-NEXT: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s32))
; CHECK-NEXT: [[S_LOAD_DWORDX3_IMM:%[0-9]+]]:sgpr_96 = S_LOAD_DWORDX3_IMM [[DEF]], 4, 0 :: (dereferenceable invariant load (s96), align 16)
%0:sgpr_64 = IMPLICIT_DEF
%1:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s32))
%2:sgpr_96 = S_LOAD_DWORDX3_IMM %0:sgpr_64, 4, 0 :: (dereferenceable invariant load (s96))
...
---
name: merge_s_load_x1_x1_x1
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x1_x1_x1
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %4:sreg_64_xexec = S_LOAD_DWORDX2_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s64), align 4)
; GFX10-NEXT: [[COPY:%[0-9]+]]:sreg_32_xm0_xexec = COPY %4.sub0
; GFX10-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed %4.sub1
; GFX10-NEXT: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[DEF]], 8, 0 :: (dereferenceable invariant load (s32))
;
; GFX11-LABEL: name: merge_s_load_x1_x1_x1
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX2_IMM:%[0-9]+]]:sreg_64_xexec = S_LOAD_DWORDX2_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s64), align 4)
; GFX11-NEXT: [[COPY:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[S_LOAD_DWORDX2_IMM]].sub0
; GFX11-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_LOAD_DWORDX2_IMM]].sub1
; GFX11-NEXT: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[DEF]], 8, 0 :: (dereferenceable invariant load (s32))
;
; GFX12-LABEL: name: merge_s_load_x1_x1_x1
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX3_IMM:%[0-9]+]]:sgpr_96 = S_LOAD_DWORDX3_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s96), align 4)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sreg_64_xexec = COPY [[S_LOAD_DWORDX3_IMM]].sub0_sub1
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_LOAD_DWORDX3_IMM]].sub2
; GFX12-NEXT: [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY]].sub0
; GFX12-NEXT: [[COPY3:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY]].sub1
%0:sgpr_64 = IMPLICIT_DEF
%1:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s32))
%2:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 4, 0 :: (dereferenceable invariant load (s32))
%3:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 8, 0 :: (dereferenceable invariant load (s32))
...
---
name: merge_s_load_x1_x1_x1_x1
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x1_x1_x1_x1
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %7:sgpr_128 = S_LOAD_DWORDX4_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s128), align 4)
; GFX10-NEXT: [[COPY:%[0-9]+]]:sreg_64_xexec = COPY %7.sub0_sub1
; GFX10-NEXT: [[COPY1:%[0-9]+]]:sreg_64_xexec = COPY killed %7.sub2_sub3
; GFX10-NEXT: [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY]].sub0
; GFX10-NEXT: [[COPY3:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY]].sub1
; GFX10-NEXT: [[COPY4:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY1]].sub0
; GFX10-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY1]].sub1
;
; GFX11-LABEL: name: merge_s_load_x1_x1_x1_x1
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_LOAD_DWORDX4_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s128), align 4)
; GFX11-NEXT: [[COPY:%[0-9]+]]:sreg_64_xexec = COPY [[S_LOAD_DWORDX4_IMM]].sub0_sub1
; GFX11-NEXT: [[COPY1:%[0-9]+]]:sreg_64_xexec = COPY killed [[S_LOAD_DWORDX4_IMM]].sub2_sub3
; GFX11-NEXT: [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY]].sub0
; GFX11-NEXT: [[COPY3:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY]].sub1
; GFX11-NEXT: [[COPY4:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY1]].sub0
; GFX11-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY1]].sub1
;
; GFX12-LABEL: name: merge_s_load_x1_x1_x1_x1
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_LOAD_DWORDX4_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s128), align 4)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_96 = COPY [[S_LOAD_DWORDX4_IMM]].sub0_sub1_sub2
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_LOAD_DWORDX4_IMM]].sub3
; GFX12-NEXT: [[COPY2:%[0-9]+]]:sreg_64_xexec = COPY [[COPY]].sub0_sub1
; GFX12-NEXT: [[COPY3:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY]].sub2
; GFX12-NEXT: [[COPY4:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY2]].sub0
; GFX12-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY2]].sub1
%0:sgpr_64 = IMPLICIT_DEF
%1:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s32))
%2:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 4, 0 :: (dereferenceable invariant load (s32))
%3:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 8, 0 :: (dereferenceable invariant load (s32))
%4:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 12, 0 :: (dereferenceable invariant load (s32))
...
---
name: merge_s_load_x1_x1_x1_x1_x1_x1_x1_x1
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x1_x1_x1_x1_x1_x1_x1_x1
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %15:sgpr_256 = S_LOAD_DWORDX8_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 4)
; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY %15.sub0_sub1_sub2_sub3
; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed %15.sub4_sub5_sub6_sub7
; GFX10-NEXT: [[COPY2:%[0-9]+]]:sreg_64_xexec = COPY [[COPY]].sub0_sub1
; GFX10-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY killed [[COPY]].sub2_sub3
; GFX10-NEXT: [[COPY4:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY2]].sub0
; GFX10-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY2]].sub1
; GFX10-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY3]].sub0
; GFX10-NEXT: [[COPY7:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub1
; GFX10-NEXT: [[COPY8:%[0-9]+]]:sreg_64_xexec = COPY [[COPY1]].sub0_sub1
; GFX10-NEXT: [[COPY9:%[0-9]+]]:sreg_64_xexec = COPY killed [[COPY1]].sub2_sub3
; GFX10-NEXT: [[COPY10:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY8]].sub0
; GFX10-NEXT: [[COPY11:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY8]].sub1
; GFX10-NEXT: [[COPY12:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY9]].sub0
; GFX10-NEXT: [[COPY13:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY9]].sub1
;
; GFX11-LABEL: name: merge_s_load_x1_x1_x1_x1_x1_x1_x1_x1
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 4)
; GFX11-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY [[S_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
; GFX11-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed [[S_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
; GFX11-NEXT: [[COPY2:%[0-9]+]]:sreg_64_xexec = COPY [[COPY]].sub0_sub1
; GFX11-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY killed [[COPY]].sub2_sub3
; GFX11-NEXT: [[COPY4:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY2]].sub0
; GFX11-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY2]].sub1
; GFX11-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY3]].sub0
; GFX11-NEXT: [[COPY7:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub1
; GFX11-NEXT: [[COPY8:%[0-9]+]]:sreg_64_xexec = COPY [[COPY1]].sub0_sub1
; GFX11-NEXT: [[COPY9:%[0-9]+]]:sreg_64_xexec = COPY killed [[COPY1]].sub2_sub3
; GFX11-NEXT: [[COPY10:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY8]].sub0
; GFX11-NEXT: [[COPY11:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY8]].sub1
; GFX11-NEXT: [[COPY12:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY9]].sub0
; GFX11-NEXT: [[COPY13:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY9]].sub1
;
; GFX12-LABEL: name: merge_s_load_x1_x1_x1_x1_x1_x1_x1_x1
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 4)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY [[S_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed [[S_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_96 = COPY [[COPY]].sub0_sub1_sub2
; GFX12-NEXT: [[COPY3:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY]].sub3
; GFX12-NEXT: [[COPY4:%[0-9]+]]:sreg_64_xexec = COPY [[COPY2]].sub0_sub1
; GFX12-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY2]].sub2
; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY4]].sub0
; GFX12-NEXT: [[COPY7:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY4]].sub1
; GFX12-NEXT: [[COPY8:%[0-9]+]]:sgpr_96 = COPY [[COPY1]].sub0_sub1_sub2
; GFX12-NEXT: [[COPY9:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY1]].sub3
; GFX12-NEXT: [[COPY10:%[0-9]+]]:sreg_64_xexec = COPY [[COPY8]].sub0_sub1
; GFX12-NEXT: [[COPY11:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY8]].sub2
; GFX12-NEXT: [[COPY12:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY10]].sub0
; GFX12-NEXT: [[COPY13:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY10]].sub1
%0:sgpr_64 = IMPLICIT_DEF
%1:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s32))
%2:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 4, 0 :: (dereferenceable invariant load (s32))
%3:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 8, 0 :: (dereferenceable invariant load (s32))
%4:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 12, 0 :: (dereferenceable invariant load (s32))
%5:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 16, 0 :: (dereferenceable invariant load (s32))
%6:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 20, 0 :: (dereferenceable invariant load (s32))
%7:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 24, 0 :: (dereferenceable invariant load (s32))
%8:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 28, 0 :: (dereferenceable invariant load (s32))
...
---
name: merge_s_load_x2_x1
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x2_x1
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: [[S_LOAD_DWORDX2_IMM:%[0-9]+]]:sgpr_64 = S_LOAD_DWORDX2_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s64))
; GFX10-NEXT: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[DEF]], 8, 0 :: (dereferenceable invariant load (s32))
;
; GFX11-LABEL: name: merge_s_load_x2_x1
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX2_IMM:%[0-9]+]]:sgpr_64 = S_LOAD_DWORDX2_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s64))
; GFX11-NEXT: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[DEF]], 8, 0 :: (dereferenceable invariant load (s32))
;
; GFX12-LABEL: name: merge_s_load_x2_x1
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX3_IMM:%[0-9]+]]:sgpr_96 = S_LOAD_DWORDX3_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s96), align 8)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_64 = COPY [[S_LOAD_DWORDX3_IMM]].sub0_sub1
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_LOAD_DWORDX3_IMM]].sub2
%0:sgpr_64 = IMPLICIT_DEF
%1:sgpr_64 = S_LOAD_DWORDX2_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s64))
%2:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 8, 0 :: (dereferenceable invariant load (s32))
...
---
name: merge_s_load_x2_x2
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x2_x2
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %3:sgpr_128 = S_LOAD_DWORDX4_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s128), align 8)
; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_64 = COPY %3.sub0_sub1
; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_64 = COPY killed %3.sub2_sub3
;
; GFX11-LABEL: name: merge_s_load_x2_x2
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_LOAD_DWORDX4_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s128), align 8)
; GFX11-NEXT: [[COPY:%[0-9]+]]:sgpr_64 = COPY [[S_LOAD_DWORDX4_IMM]].sub0_sub1
; GFX11-NEXT: [[COPY1:%[0-9]+]]:sgpr_64 = COPY killed [[S_LOAD_DWORDX4_IMM]].sub2_sub3
;
; GFX12-LABEL: name: merge_s_load_x2_x2
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_LOAD_DWORDX4_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s128), align 8)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_64 = COPY [[S_LOAD_DWORDX4_IMM]].sub0_sub1
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_64 = COPY killed [[S_LOAD_DWORDX4_IMM]].sub2_sub3
%0:sgpr_64 = IMPLICIT_DEF
%1:sgpr_64 = S_LOAD_DWORDX2_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s64))
%2:sgpr_64 = S_LOAD_DWORDX2_IMM %0:sgpr_64, 8, 0 :: (dereferenceable invariant load (s64))
...
---
name: merge_s_load_x2_x2_x2_x2
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x2_x2_x2_x2
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %7:sgpr_256 = S_LOAD_DWORDX8_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY %7.sub0_sub1_sub2_sub3
; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed %7.sub4_sub5_sub6_sub7
; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_64 = COPY [[COPY]].sub0_sub1
; GFX10-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY killed [[COPY]].sub2_sub3
; GFX10-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
; GFX10-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
;
; GFX11-LABEL: name: merge_s_load_x2_x2_x2_x2
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
; GFX11-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY [[S_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
; GFX11-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed [[S_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
; GFX11-NEXT: [[COPY2:%[0-9]+]]:sgpr_64 = COPY [[COPY]].sub0_sub1
; GFX11-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY killed [[COPY]].sub2_sub3
; GFX11-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
; GFX11-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
;
; GFX12-LABEL: name: merge_s_load_x2_x2_x2_x2
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY [[S_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed [[S_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_64 = COPY [[COPY]].sub0_sub1
; GFX12-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY killed [[COPY]].sub2_sub3
; GFX12-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
; GFX12-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
%0:sgpr_64 = IMPLICIT_DEF
%1:sgpr_64 = S_LOAD_DWORDX2_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s64))
%2:sgpr_64 = S_LOAD_DWORDX2_IMM %0:sgpr_64, 8, 0 :: (dereferenceable invariant load (s64))
%3:sgpr_64 = S_LOAD_DWORDX2_IMM %0:sgpr_64, 16, 0 :: (dereferenceable invariant load (s64))
%4:sgpr_64 = S_LOAD_DWORDX2_IMM %0:sgpr_64, 24, 0 :: (dereferenceable invariant load (s64))
...
---
name: merge_s_load_x3_x1
body: |
bb.0:
; CHECK-LABEL: name: merge_s_load_x3_x1
; CHECK: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; CHECK-NEXT: [[S_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_LOAD_DWORDX4_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s128))
; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_96 = COPY [[S_LOAD_DWORDX4_IMM]].sub0_sub1_sub2
; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_LOAD_DWORDX4_IMM]].sub3
%0:sgpr_64 = IMPLICIT_DEF
%1:sgpr_96 = S_LOAD_DWORDX3_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s96))
%2:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 12, 0 :: (dereferenceable invariant load (s32))
...
---
name: merge_s_load_x4_x4
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x4_x4
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %3:sgpr_256 = S_LOAD_DWORDX8_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY %3.sub0_sub1_sub2_sub3
; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed %3.sub4_sub5_sub6_sub7
;
; GFX11-LABEL: name: merge_s_load_x4_x4
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
; GFX11-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY [[S_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
; GFX11-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed [[S_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
;
; GFX12-LABEL: name: merge_s_load_x4_x4
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY [[S_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed [[S_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
%0:sgpr_64 = IMPLICIT_DEF
%1:sgpr_128 = S_LOAD_DWORDX4_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s128))
%2:sgpr_128 = S_LOAD_DWORDX4_IMM %0:sgpr_64, 16, 0 :: (dereferenceable invariant load (s128))
...
# The constrained multi-dword scalar load merge tests.
---
name: merge_s_load_x1_x2ec
body: |
bb.0:
; CHECK-LABEL: name: merge_s_load_x1_x2ec
; CHECK: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; CHECK-NEXT: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s32))
; CHECK-NEXT: early-clobber %2:sgpr_64 = S_LOAD_DWORDX2_IMM_ec [[DEF]], 4, 0 :: (dereferenceable invariant load (s64))
%0:sgpr_64 = IMPLICIT_DEF
%1:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s32))
early-clobber %2:sgpr_64 = S_LOAD_DWORDX2_IMM_ec %0:sgpr_64, 4, 0 :: (dereferenceable invariant load (s64))
...
---
name: merge_s_load_x1_x3ec
body: |
bb.0:
; CHECK-LABEL: name: merge_s_load_x1_x3ec
; CHECK: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; CHECK-NEXT: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s32))
; CHECK-NEXT: early-clobber %2:sgpr_96 = S_LOAD_DWORDX3_IMM_ec [[DEF]], 4, 0 :: (dereferenceable invariant load (s96), align 16)
%0:sgpr_64 = IMPLICIT_DEF
%1:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s32))
early-clobber %2:sgpr_96 = S_LOAD_DWORDX3_IMM_ec %0:sgpr_64, 4, 0 :: (dereferenceable invariant load (s96))
...
---
name: merge_s_load_x2ec_x1
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x2ec_x1
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %1:sgpr_64 = S_LOAD_DWORDX2_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s64))
; GFX10-NEXT: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[DEF]], 8, 0 :: (dereferenceable invariant load (s32))
;
; GFX11-LABEL: name: merge_s_load_x2ec_x1
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: early-clobber %1:sgpr_64 = S_LOAD_DWORDX2_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s64))
; GFX11-NEXT: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[DEF]], 8, 0 :: (dereferenceable invariant load (s32))
;
; GFX12-LABEL: name: merge_s_load_x2ec_x1
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX3_IMM:%[0-9]+]]:sgpr_96 = S_LOAD_DWORDX3_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s96), align 8)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_64 = COPY [[S_LOAD_DWORDX3_IMM]].sub0_sub1
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_LOAD_DWORDX3_IMM]].sub2
%0:sgpr_64 = IMPLICIT_DEF
early-clobber %1:sgpr_64 = S_LOAD_DWORDX2_IMM_ec %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s64))
%2:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 8, 0 :: (dereferenceable invariant load (s32))
...
---
name: merge_s_load_x2ec_x2ec
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x2ec_x2ec
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %3:sgpr_128 = S_LOAD_DWORDX4_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s128), align 8)
; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_64 = COPY %3.sub0_sub1
; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_64 = COPY killed %3.sub2_sub3
;
; GFX11-LABEL: name: merge_s_load_x2ec_x2ec
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_LOAD_DWORDX4_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s128), align 8)
; GFX11-NEXT: [[COPY:%[0-9]+]]:sgpr_64 = COPY [[S_LOAD_DWORDX4_IMM]].sub0_sub1
; GFX11-NEXT: [[COPY1:%[0-9]+]]:sgpr_64 = COPY killed [[S_LOAD_DWORDX4_IMM]].sub2_sub3
;
; GFX12-LABEL: name: merge_s_load_x2ec_x2ec
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_LOAD_DWORDX4_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s128), align 8)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_64 = COPY [[S_LOAD_DWORDX4_IMM]].sub0_sub1
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_64 = COPY killed [[S_LOAD_DWORDX4_IMM]].sub2_sub3
%0:sgpr_64 = IMPLICIT_DEF
early-clobber %1:sgpr_64 = S_LOAD_DWORDX2_IMM_ec %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s64))
early-clobber %2:sgpr_64 = S_LOAD_DWORDX2_IMM_ec %0:sgpr_64, 8, 0 :: (dereferenceable invariant load (s64))
...
---
name: merge_s_load_x2ec_x2ec_x2ec_x2ec
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x2ec_x2ec_x2ec_x2ec
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %7:sgpr_256 = S_LOAD_DWORDX8_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY %7.sub0_sub1_sub2_sub3
; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed %7.sub4_sub5_sub6_sub7
; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_64 = COPY [[COPY]].sub0_sub1
; GFX10-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY killed [[COPY]].sub2_sub3
; GFX10-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
; GFX10-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
;
; GFX11-LABEL: name: merge_s_load_x2ec_x2ec_x2ec_x2ec
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
; GFX11-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY [[S_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
; GFX11-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed [[S_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
; GFX11-NEXT: [[COPY2:%[0-9]+]]:sgpr_64 = COPY [[COPY]].sub0_sub1
; GFX11-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY killed [[COPY]].sub2_sub3
; GFX11-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
; GFX11-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
;
; GFX12-LABEL: name: merge_s_load_x2ec_x2ec_x2ec_x2ec
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY [[S_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed [[S_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_64 = COPY [[COPY]].sub0_sub1
; GFX12-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY killed [[COPY]].sub2_sub3
; GFX12-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
; GFX12-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
%0:sgpr_64 = IMPLICIT_DEF
early-clobber %1:sgpr_64 = S_LOAD_DWORDX2_IMM_ec %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s64))
early-clobber %2:sgpr_64 = S_LOAD_DWORDX2_IMM_ec %0:sgpr_64, 8, 0 :: (dereferenceable invariant load (s64))
early-clobber %3:sgpr_64 = S_LOAD_DWORDX2_IMM_ec %0:sgpr_64, 16, 0 :: (dereferenceable invariant load (s64))
early-clobber %4:sgpr_64 = S_LOAD_DWORDX2_IMM_ec %0:sgpr_64, 24, 0 :: (dereferenceable invariant load (s64))
...
---
name: merge_s_load_x3ec_x1
body: |
bb.0:
; CHECK-LABEL: name: merge_s_load_x3ec_x1
; CHECK: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; CHECK-NEXT: [[S_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_LOAD_DWORDX4_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s128))
; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_96 = COPY [[S_LOAD_DWORDX4_IMM]].sub0_sub1_sub2
; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_LOAD_DWORDX4_IMM]].sub3
%0:sgpr_64 = IMPLICIT_DEF
early-clobber %1:sgpr_96 = S_LOAD_DWORDX3_IMM_ec %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s96))
%2:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0:sgpr_64, 12, 0 :: (dereferenceable invariant load (s32))
...
---
name: merge_s_load_x4ec_x4ec
body: |
bb.0:
; GFX10-LABEL: name: merge_s_load_x4ec_x4ec
; GFX10: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX10-NEXT: early-clobber %3:sgpr_256 = S_LOAD_DWORDX8_IMM_ec [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY %3.sub0_sub1_sub2_sub3
; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed %3.sub4_sub5_sub6_sub7
;
; GFX11-LABEL: name: merge_s_load_x4ec_x4ec
; GFX11: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX11-NEXT: [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
; GFX11-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY [[S_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
; GFX11-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed [[S_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
;
; GFX12-LABEL: name: merge_s_load_x4ec_x4ec
; GFX12: [[DEF:%[0-9]+]]:sgpr_64 = IMPLICIT_DEF
; GFX12-NEXT: [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[DEF]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY [[S_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY killed [[S_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
%0:sgpr_64 = IMPLICIT_DEF
early-clobber %1:sgpr_128 = S_LOAD_DWORDX4_IMM_ec %0:sgpr_64, 0, 0 :: (dereferenceable invariant load (s128))
early-clobber %2:sgpr_128 = S_LOAD_DWORDX4_IMM_ec %0:sgpr_64, 16, 0 :: (dereferenceable invariant load (s128))
...
Codebase Browser
llvm