llvm/test/CodeGen/AMDGPU/merge-sbuffer-load.mir

   1 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
   2 # RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs -run-pass si-load-store-opt -o - %s | FileCheck %s -check-prefixes=CHECK,GFX10
   3 # RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs -run-pass si-load-store-opt -o - %s | FileCheck %s -check-prefixes=CHECK,GFX12
   4
   5 ---
   6 name: merge_s_buffer_load_x2
   7 tracksRegLiveness: true
   8 body:             |
   9   bb.0:
  10     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
  11
  12     ; GFX10-LABEL: name: merge_s_buffer_load_x2
  13     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
  14     ; GFX10-NEXT: {{  $}}
  15     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
  16     ; GFX10-NEXT: early-clobber %3:sreg_64_xexec = S_BUFFER_LOAD_DWORDX2_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s64), align 4)
  17     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY %3.sub0
  18     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed %3.sub1
  19     ; GFX10-NEXT: S_ENDPGM 0
  20     ;
  21     ; GFX12-LABEL: name: merge_s_buffer_load_x2
  22     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
  23     ; GFX12-NEXT: {{  $}}
  24     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
  25     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX2_IMM:%[0-9]+]]:sreg_64_xexec = S_BUFFER_LOAD_DWORDX2_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s64), align 4)
  26     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[S_BUFFER_LOAD_DWORDX2_IMM]].sub0
  27     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_BUFFER_LOAD_DWORDX2_IMM]].sub1
  28     ; GFX12-NEXT: S_ENDPGM 0
  29     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
  30     %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
  31     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s32))
  32
  33     S_ENDPGM 0
  34 ...
  35 ---
  36
  37 name: merge_s_buffer_load_x1_x2
  38 tracksRegLiveness: true
  39 body:             |
  40   bb.0:
  41     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
  42
  43     ; CHECK-LABEL: name: merge_s_buffer_load_x1_x2
  44     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
  45     ; CHECK-NEXT: {{  $}}
  46     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
  47     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s32))
  48     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORDX2_IMM:%[0-9]+]]:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM [[COPY]], 4, 0 :: (dereferenceable invariant load (s64))
  49     ; CHECK-NEXT: S_ENDPGM 0
  50     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
  51     %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
  52     %2:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s64))
  53
  54     S_ENDPGM 0
  55 ...
  56 ---
  57
  58 name: merge_s_buffer_load_x2_x1
  59 tracksRegLiveness: true
  60 body:             |
  61   bb.0:
  62     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
  63
  64     ; GFX10-LABEL: name: merge_s_buffer_load_x2_x1
  65     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
  66     ; GFX10-NEXT: {{  $}}
  67     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
  68     ; GFX10-NEXT: [[S_BUFFER_LOAD_DWORDX2_IMM:%[0-9]+]]:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s64))
  69     ; GFX10-NEXT: [[S_BUFFER_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM [[COPY]], 8, 0 :: (dereferenceable invariant load (s32))
  70     ; GFX10-NEXT: S_ENDPGM 0
  71     ;
  72     ; GFX12-LABEL: name: merge_s_buffer_load_x2_x1
  73     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
  74     ; GFX12-NEXT: {{  $}}
  75     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
  76     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX3_IMM:%[0-9]+]]:sgpr_96 = S_BUFFER_LOAD_DWORDX3_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s96), align 8)
  77     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_64 = COPY [[S_BUFFER_LOAD_DWORDX3_IMM]].sub0_sub1
  78     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_BUFFER_LOAD_DWORDX3_IMM]].sub2
  79     ; GFX12-NEXT: S_ENDPGM 0
  80     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
  81     %1:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s64))
  82     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s32))
  83
  84     S_ENDPGM 0
  85 ...
  86 ---
  87
  88 name: merge_s_buffer_load_x4
  89 tracksRegLiveness: true
  90 body:             |
  91   bb.0:
  92     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
  93
  94     ; GFX10-LABEL: name: merge_s_buffer_load_x4
  95     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
  96     ; GFX10-NEXT: {{  $}}
  97     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
  98     ; GFX10-NEXT: early-clobber %7:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s128), align 4)
  99     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sreg_64_xexec = COPY %7.sub0_sub1
 100     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sreg_64_xexec = COPY killed %7.sub2_sub3
 101     ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY1]].sub0
 102     ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY1]].sub1
 103     ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY2]].sub0
 104     ; GFX10-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY2]].sub1
 105     ; GFX10-NEXT: S_ENDPGM 0
 106     ;
 107     ; GFX12-LABEL: name: merge_s_buffer_load_x4
 108     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 109     ; GFX12-NEXT: {{  $}}
 110     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 111     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s128), align 4)
 112     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_96 = COPY [[S_BUFFER_LOAD_DWORDX4_IMM]].sub0_sub1_sub2
 113     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_BUFFER_LOAD_DWORDX4_IMM]].sub3
 114     ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY [[COPY1]].sub0_sub1
 115     ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY1]].sub2
 116     ; GFX12-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY3]].sub0
 117     ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub1
 118     ; GFX12-NEXT: S_ENDPGM 0
 119     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 120     %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
 121     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s32))
 122     %3:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s32))
 123     %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 12, 0 :: (dereferenceable invariant load (s32))
 124
 125     S_ENDPGM 0
 126 ...
 127 ---
 128
 129 name: merge_s_buffer_load_x1_x3
 130 tracksRegLiveness: true
 131 body:             |
 132   bb.0:
 133     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 134
 135     ; CHECK-LABEL: name: merge_s_buffer_load_x1_x3
 136     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 137     ; CHECK-NEXT: {{  $}}
 138     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 139     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s32))
 140     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORDX3_IMM:%[0-9]+]]:sgpr_96 = S_BUFFER_LOAD_DWORDX3_IMM [[COPY]], 4, 0 :: (dereferenceable invariant load (s96), align 16)
 141     ; CHECK-NEXT: S_ENDPGM 0
 142     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 143     %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
 144     %2:sgpr_96 = S_BUFFER_LOAD_DWORDX3_IMM %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s96))
 145
 146     S_ENDPGM 0
 147 ...
 148 ---
 149
 150 name: merge_s_buffer_load_x3_x1
 151 tracksRegLiveness: true
 152 body:             |
 153   bb.0:
 154     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 155
 156     ; CHECK-LABEL: name: merge_s_buffer_load_x3_x1
 157     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 158     ; CHECK-NEXT: {{  $}}
 159     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 160     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s128))
 161     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr_96 = COPY [[S_BUFFER_LOAD_DWORDX4_IMM]].sub0_sub1_sub2
 162     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_BUFFER_LOAD_DWORDX4_IMM]].sub3
 163     ; CHECK-NEXT: S_ENDPGM 0
 164     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 165     %1:sgpr_96 = S_BUFFER_LOAD_DWORDX3_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s96))
 166     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 12, 0 :: (dereferenceable invariant load (s32))
 167
 168     S_ENDPGM 0
 169 ...
 170 ---
 171
 172 name: merge_s_buffer_load_x8
 173 tracksRegLiveness: true
 174 body:             |
 175   bb.0:
 176     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 177
 178     ; GFX10-LABEL: name: merge_s_buffer_load_x8
 179     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 180     ; GFX10-NEXT: {{  $}}
 181     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 182     ; GFX10-NEXT: early-clobber %15:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 4)
 183     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %15.sub0_sub1_sub2_sub3
 184     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %15.sub4_sub5_sub6_sub7
 185     ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY [[COPY1]].sub0_sub1
 186     ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sreg_64_xexec = COPY killed [[COPY1]].sub2_sub3
 187     ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY3]].sub0
 188     ; GFX10-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub1
 189     ; GFX10-NEXT: [[COPY7:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY4]].sub0
 190     ; GFX10-NEXT: [[COPY8:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY4]].sub1
 191     ; GFX10-NEXT: [[COPY9:%[0-9]+]]:sreg_64_xexec = COPY [[COPY2]].sub0_sub1
 192     ; GFX10-NEXT: [[COPY10:%[0-9]+]]:sreg_64_xexec = COPY killed [[COPY2]].sub2_sub3
 193     ; GFX10-NEXT: [[COPY11:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY9]].sub0
 194     ; GFX10-NEXT: [[COPY12:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY9]].sub1
 195     ; GFX10-NEXT: [[COPY13:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY10]].sub0
 196     ; GFX10-NEXT: [[COPY14:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY10]].sub1
 197     ; GFX10-NEXT: S_ENDPGM 0
 198     ;
 199     ; GFX12-LABEL: name: merge_s_buffer_load_x8
 200     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 201     ; GFX12-NEXT: {{  $}}
 202     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 203     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 4)
 204     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 205     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 206     ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sgpr_96 = COPY [[COPY1]].sub0_sub1_sub2
 207     ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY1]].sub3
 208     ; GFX12-NEXT: [[COPY5:%[0-9]+]]:sreg_64_xexec = COPY [[COPY3]].sub0_sub1
 209     ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub2
 210     ; GFX12-NEXT: [[COPY7:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY5]].sub0
 211     ; GFX12-NEXT: [[COPY8:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY5]].sub1
 212     ; GFX12-NEXT: [[COPY9:%[0-9]+]]:sgpr_96 = COPY [[COPY2]].sub0_sub1_sub2
 213     ; GFX12-NEXT: [[COPY10:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY2]].sub3
 214     ; GFX12-NEXT: [[COPY11:%[0-9]+]]:sreg_64_xexec = COPY [[COPY9]].sub0_sub1
 215     ; GFX12-NEXT: [[COPY12:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY9]].sub2
 216     ; GFX12-NEXT: [[COPY13:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY11]].sub0
 217     ; GFX12-NEXT: [[COPY14:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY11]].sub1
 218     ; GFX12-NEXT: S_ENDPGM 0
 219     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 220     %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
 221     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s32))
 222     %3:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s32))
 223     %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 12, 0 :: (dereferenceable invariant load (s32))
 224     %5:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s32))
 225     %6:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 20, 0 :: (dereferenceable invariant load (s32))
 226     %7:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s32))
 227     %8:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 28, 0 :: (dereferenceable invariant load (s32))
 228
 229     S_ENDPGM 0
 230 ...
 231 ---
 232
 233 name: merge_s_buffer_load_x8_reordered
 234 tracksRegLiveness: true
 235 body:             |
 236   bb.0:
 237     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 238
 239     ; GFX10-LABEL: name: merge_s_buffer_load_x8_reordered
 240     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 241     ; GFX10-NEXT: {{  $}}
 242     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 243     ; GFX10-NEXT: early-clobber %15:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 4)
 244     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %15.sub4_sub5_sub6_sub7
 245     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %15.sub0_sub1_sub2_sub3
 246     ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY [[COPY1]].sub0_sub1
 247     ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sreg_64_xexec = COPY killed [[COPY1]].sub2_sub3
 248     ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY3]].sub1
 249     ; GFX10-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub0
 250     ; GFX10-NEXT: [[COPY7:%[0-9]+]]:sreg_64_xexec = COPY [[COPY2]].sub0_sub1
 251     ; GFX10-NEXT: [[COPY8:%[0-9]+]]:sreg_64_xexec = COPY killed [[COPY2]].sub2_sub3
 252     ; GFX10-NEXT: [[COPY9:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY7]].sub1
 253     ; GFX10-NEXT: [[COPY10:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY7]].sub0
 254     ; GFX10-NEXT: [[COPY11:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY4]].sub1
 255     ; GFX10-NEXT: [[COPY12:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY4]].sub0
 256     ; GFX10-NEXT: [[COPY13:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY8]].sub1
 257     ; GFX10-NEXT: [[COPY14:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY8]].sub0
 258     ; GFX10-NEXT: S_ENDPGM 0
 259     ;
 260     ; GFX12-LABEL: name: merge_s_buffer_load_x8_reordered
 261     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 262     ; GFX12-NEXT: {{  $}}
 263     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 264     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 4)
 265     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 266     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 267     ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sgpr_96 = COPY [[COPY1]].sub0_sub1_sub2
 268     ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY1]].sub3
 269     ; GFX12-NEXT: [[COPY5:%[0-9]+]]:sreg_64_xexec = COPY [[COPY3]].sub0_sub1
 270     ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub2
 271     ; GFX12-NEXT: [[COPY7:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY5]].sub1
 272     ; GFX12-NEXT: [[COPY8:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY5]].sub0
 273     ; GFX12-NEXT: [[COPY9:%[0-9]+]]:sgpr_96 = COPY [[COPY2]].sub0_sub1_sub2
 274     ; GFX12-NEXT: [[COPY10:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY2]].sub3
 275     ; GFX12-NEXT: [[COPY11:%[0-9]+]]:sreg_64_xexec = COPY [[COPY9]].sub0_sub1
 276     ; GFX12-NEXT: [[COPY12:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY9]].sub2
 277     ; GFX12-NEXT: [[COPY13:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY11]].sub1
 278     ; GFX12-NEXT: [[COPY14:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY11]].sub0
 279     ; GFX12-NEXT: S_ENDPGM 0
 280     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 281     %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 20, 0 :: (dereferenceable invariant load (s32))
 282     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s32))
 283     %3:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
 284     %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 28, 0 :: (dereferenceable invariant load (s32))
 285     %5:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 12, 0 :: (dereferenceable invariant load (s32))
 286     %6:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s32))
 287     %7:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s32))
 288     %8:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s32))
 289
 290     S_ENDPGM 0
 291 ...
 292 ---
 293
 294 name: merge_s_buffer_load_x8_out_of_x2
 295 tracksRegLiveness: true
 296 body:             |
 297   bb.0:
 298     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 299
 300     ; GFX10-LABEL: name: merge_s_buffer_load_x8_out_of_x2
 301     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 302     ; GFX10-NEXT: {{  $}}
 303     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 304     ; GFX10-NEXT: early-clobber %7:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
 305     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %7.sub4_sub5_sub6_sub7
 306     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %7.sub0_sub1_sub2_sub3
 307     ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
 308     ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
 309     ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY [[COPY2]].sub2_sub3
 310     ; GFX10-NEXT: [[COPY6:%[0-9]+]]:sgpr_64 = COPY killed [[COPY2]].sub0_sub1
 311     ; GFX10-NEXT: S_ENDPGM 0
 312     ;
 313     ; GFX12-LABEL: name: merge_s_buffer_load_x8_out_of_x2
 314     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 315     ; GFX12-NEXT: {{  $}}
 316     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 317     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
 318     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 319     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 320     ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
 321     ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
 322     ; GFX12-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY [[COPY2]].sub2_sub3
 323     ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sgpr_64 = COPY killed [[COPY2]].sub0_sub1
 324     ; GFX12-NEXT: S_ENDPGM 0
 325     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 326     %1:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s64))
 327     %2:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s64))
 328     %3:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s64))
 329     %4:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s64))
 330
 331     S_ENDPGM 0
 332 ...
 333 ---
 334
 335 name: merge_s_buffer_load_x8_out_of_x4
 336 tracksRegLiveness: true
 337 body:             |
 338   bb.0:
 339     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 340
 341     ; GFX10-LABEL: name: merge_s_buffer_load_x8_out_of_x4
 342     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 343     ; GFX10-NEXT: {{  $}}
 344     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 345     ; GFX10-NEXT: early-clobber %3:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 346     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %3.sub0_sub1_sub2_sub3
 347     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %3.sub4_sub5_sub6_sub7
 348     ; GFX10-NEXT: S_ENDPGM 0
 349     ;
 350     ; GFX12-LABEL: name: merge_s_buffer_load_x8_out_of_x4
 351     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 352     ; GFX12-NEXT: {{  $}}
 353     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 354     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 355     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 356     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 357     ; GFX12-NEXT: S_ENDPGM 0
 358     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 359     %1:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s128))
 360     %2:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s128))
 361
 362     S_ENDPGM 0
 363 ...
 364 ---
 365
 366 name: merge_s_buffer_load_x8_mixed
 367 tracksRegLiveness: true
 368 body:             |
 369   bb.0:
 370     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 371
 372     ; GFX10-LABEL: name: merge_s_buffer_load_x8_mixed
 373     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 374     ; GFX10-NEXT: {{  $}}
 375     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 376     ; GFX10-NEXT: early-clobber %7:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 377     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %7.sub0_sub1_sub2_sub3
 378     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %7.sub4_sub5_sub6_sub7
 379     ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY [[COPY2]].sub0_sub1
 380     ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY killed [[COPY2]].sub2_sub3
 381     ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY3]].sub0
 382     ; GFX10-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub1
 383     ; GFX10-NEXT: S_ENDPGM 0
 384     ;
 385     ; GFX12-LABEL: name: merge_s_buffer_load_x8_mixed
 386     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 387     ; GFX12-NEXT: {{  $}}
 388     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 389     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 390     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 391     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 392     ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY [[COPY2]].sub0_sub1
 393     ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY killed [[COPY2]].sub2_sub3
 394     ; GFX12-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY3]].sub0
 395     ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub1
 396     ; GFX12-NEXT: S_ENDPGM 0
 397     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 398     %1:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s128))
 399     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s32))
 400     %3:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s64))
 401     %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 20, 0 :: (dereferenceable invariant load (s32))
 402
 403     S_ENDPGM 0
 404 ...
 405 ---
 406
 407 name: merge_s_buffer_load_sgpr_imm
 408 tracksRegLiveness: true
 409 body:             |
 410   bb.0:
 411     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
 412
 413     ; GFX10-LABEL: name: merge_s_buffer_load_sgpr_imm
 414     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
 415     ; GFX10-NEXT: {{  $}}
 416     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 417     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr4
 418     ; GFX10-NEXT: early-clobber %8:sgpr_128 = S_BUFFER_LOAD_DWORDX4_SGPR_IMM_ec [[COPY]], [[COPY1]], 0, 0 :: (dereferenceable invariant load (s128), align 4)
 419     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sreg_64_xexec = COPY %8.sub0_sub1
 420     ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY killed %8.sub2_sub3
 421     ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY2]].sub0
 422     ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY2]].sub1
 423     ; GFX10-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY3]].sub0
 424     ; GFX10-NEXT: [[COPY7:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub1
 425     ; GFX10-NEXT: S_ENDPGM 0
 426     ;
 427     ; GFX12-LABEL: name: merge_s_buffer_load_sgpr_imm
 428     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
 429     ; GFX12-NEXT: {{  $}}
 430     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 431     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr4
 432     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX4_SGPR_IMM:%[0-9]+]]:sgpr_128 = S_BUFFER_LOAD_DWORDX4_SGPR_IMM [[COPY]], [[COPY1]], 0, 0 :: (dereferenceable invariant load (s128), align 4)
 433     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_96 = COPY [[S_BUFFER_LOAD_DWORDX4_SGPR_IMM]].sub0_sub1_sub2
 434     ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_BUFFER_LOAD_DWORDX4_SGPR_IMM]].sub3
 435     ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sreg_64_xexec = COPY [[COPY2]].sub0_sub1
 436     ; GFX12-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY2]].sub2
 437     ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY4]].sub0
 438     ; GFX12-NEXT: [[COPY7:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY4]].sub1
 439     ; GFX12-NEXT: S_ENDPGM 0
 440     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 441     %1:sreg_32 = COPY $sgpr4
 442     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 0, 0 :: (dereferenceable invariant load (s32))
 443     %3:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 4, 0 :: (dereferenceable invariant load (s32))
 444     %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 8, 0 :: (dereferenceable invariant load (s32))
 445     %5:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 12, 0 :: (dereferenceable invariant load (s32))
 446
 447     S_ENDPGM 0
 448 ...
 449 ---
 450
 451 name: no_merge_for_different_soffsets
 452 tracksRegLiveness: true
 453 body:             |
 454   bb.0:
 455     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $sgpr5
 456
 457     ; CHECK-LABEL: name: no_merge_for_different_soffsets
 458     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $sgpr5
 459     ; CHECK-NEXT: {{  $}}
 460     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 461     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr4
 462     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr5
 463     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORD_SGPR_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM [[COPY]], [[COPY1]], 4, 0 :: (dereferenceable invariant load (s32))
 464     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORD_SGPR_IMM1:%[0-9]+]]:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM [[COPY]], [[COPY2]], 8, 0 :: (dereferenceable invariant load (s32))
 465     ; CHECK-NEXT: S_ENDPGM 0
 466     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 467     %1:sreg_32 = COPY $sgpr4
 468     %2:sreg_32 = COPY $sgpr5
 469     %3:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 4, 0 :: (dereferenceable invariant load (s32))
 470     %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %2:sreg_32, 8, 0 :: (dereferenceable invariant load (s32))
 471
 472     S_ENDPGM 0
 473 ...
 474 ---
 475
 476 name: no_merge_for_non_adjacent_offsets
 477 tracksRegLiveness: true
 478 body:             |
 479   bb.0:
 480     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
 481
 482     ; CHECK-LABEL: name: no_merge_for_non_adjacent_offsets
 483     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
 484     ; CHECK-NEXT: {{  $}}
 485     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 486     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr4
 487     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORD_SGPR_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM [[COPY]], [[COPY1]], 4, 0 :: (dereferenceable invariant load (s32))
 488     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORD_SGPR_IMM1:%[0-9]+]]:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM [[COPY]], [[COPY1]], 12, 0 :: (dereferenceable invariant load (s32))
 489     ; CHECK-NEXT: S_ENDPGM 0
 490     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 491     %1:sreg_32 = COPY $sgpr4
 492     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 4, 0 :: (dereferenceable invariant load (s32))
 493     %3:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_SGPR_IMM %0:sgpr_128, %1:sreg_32, 12, 0 :: (dereferenceable invariant load (s32))
 494
 495     S_ENDPGM 0
 496 ...
 497
 498 # The constrained multi-dword buffer load merge tests.
 499
 500 ---
 501 name: merge_s_buffer_load_x1_x2ec
 502 tracksRegLiveness: true
 503 body:             |
 504   bb.0:
 505     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 506
 507     ; CHECK-LABEL: name: merge_s_buffer_load_x1_x2ec
 508     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 509     ; CHECK-NEXT: {{  $}}
 510     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 511     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s32))
 512     ; CHECK-NEXT: early-clobber %2:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec [[COPY]], 4, 0 :: (dereferenceable invariant load (s64))
 513     ; CHECK-NEXT: S_ENDPGM 0
 514     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 515     %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
 516     early-clobber %2:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s64))
 517
 518     S_ENDPGM 0
 519 ...
 520 ---
 521
 522 name: merge_s_buffer_load_x2ec_x1
 523 tracksRegLiveness: true
 524 body:             |
 525   bb.0:
 526     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 527
 528     ; GFX10-LABEL: name: merge_s_buffer_load_x2ec_x1
 529     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 530     ; GFX10-NEXT: {{  $}}
 531     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 532     ; GFX10-NEXT: early-clobber %1:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s64))
 533     ; GFX10-NEXT: [[S_BUFFER_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM [[COPY]], 8, 0 :: (dereferenceable invariant load (s32))
 534     ; GFX10-NEXT: S_ENDPGM 0
 535     ;
 536     ; GFX12-LABEL: name: merge_s_buffer_load_x2ec_x1
 537     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 538     ; GFX12-NEXT: {{  $}}
 539     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 540     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX3_IMM:%[0-9]+]]:sgpr_96 = S_BUFFER_LOAD_DWORDX3_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s96), align 8)
 541     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_64 = COPY [[S_BUFFER_LOAD_DWORDX3_IMM]].sub0_sub1
 542     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_BUFFER_LOAD_DWORDX3_IMM]].sub2
 543     ; GFX12-NEXT: S_ENDPGM 0
 544     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 545     early-clobber %1:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s64))
 546     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s32))
 547
 548     S_ENDPGM 0
 549 ...
 550 ---
 551
 552 name: merge_s_buffer_load_x1_x3ec
 553 tracksRegLiveness: true
 554 body:             |
 555   bb.0:
 556     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 557
 558     ; CHECK-LABEL: name: merge_s_buffer_load_x1_x3ec
 559     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 560     ; CHECK-NEXT: {{  $}}
 561     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 562     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s32))
 563     ; CHECK-NEXT: early-clobber %2:sgpr_96 = S_BUFFER_LOAD_DWORDX3_IMM_ec [[COPY]], 4, 0 :: (dereferenceable invariant load (s96), align 16)
 564     ; CHECK-NEXT: S_ENDPGM 0
 565     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 566     %1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s32))
 567     early-clobber %2:sgpr_96 = S_BUFFER_LOAD_DWORDX3_IMM_ec %0:sgpr_128, 4, 0 :: (dereferenceable invariant load (s96))
 568
 569     S_ENDPGM 0
 570 ...
 571 ---
 572
 573 name: merge_s_buffer_load_x3ec_x1
 574 tracksRegLiveness: true
 575 body:             |
 576   bb.0:
 577     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 578
 579     ; CHECK-LABEL: name: merge_s_buffer_load_x3ec_x1
 580     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 581     ; CHECK-NEXT: {{  $}}
 582     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 583     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s128))
 584     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr_96 = COPY [[S_BUFFER_LOAD_DWORDX4_IMM]].sub0_sub1_sub2
 585     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[S_BUFFER_LOAD_DWORDX4_IMM]].sub3
 586     ; CHECK-NEXT: S_ENDPGM 0
 587     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 588     early-clobber %1:sgpr_96 = S_BUFFER_LOAD_DWORDX3_IMM_ec %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s96))
 589     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 12, 0 :: (dereferenceable invariant load (s32))
 590
 591     S_ENDPGM 0
 592 ...
 593 ---
 594
 595 name: merge_s_buffer_load_x8_out_of_x2ec_reordered
 596 tracksRegLiveness: true
 597 body:             |
 598   bb.0:
 599     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 600
 601     ; GFX10-LABEL: name: merge_s_buffer_load_x8_out_of_x2ec_reordered
 602     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 603     ; GFX10-NEXT: {{  $}}
 604     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 605     ; GFX10-NEXT: early-clobber %7:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
 606     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %7.sub4_sub5_sub6_sub7
 607     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %7.sub0_sub1_sub2_sub3
 608     ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
 609     ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
 610     ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY [[COPY2]].sub2_sub3
 611     ; GFX10-NEXT: [[COPY6:%[0-9]+]]:sgpr_64 = COPY killed [[COPY2]].sub0_sub1
 612     ; GFX10-NEXT: S_ENDPGM 0
 613     ;
 614     ; GFX12-LABEL: name: merge_s_buffer_load_x8_out_of_x2ec_reordered
 615     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 616     ; GFX12-NEXT: {{  $}}
 617     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 618     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
 619     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 620     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 621     ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
 622     ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
 623     ; GFX12-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY [[COPY2]].sub2_sub3
 624     ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sgpr_64 = COPY killed [[COPY2]].sub0_sub1
 625     ; GFX12-NEXT: S_ENDPGM 0
 626     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 627     early-clobber %1:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s64))
 628     early-clobber %2:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s64))
 629     early-clobber %3:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s64))
 630     early-clobber %4:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s64))
 631
 632     S_ENDPGM 0
 633 ...
 634 ---
 635
 636 name: merge_s_buffer_load_x8_out_of_x2ec_x2
 637 tracksRegLiveness: true
 638 body:             |
 639   bb.0:
 640     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 641
 642     ; GFX10-LABEL: name: merge_s_buffer_load_x8_out_of_x2ec_x2
 643     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 644     ; GFX10-NEXT: {{  $}}
 645     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 646     ; GFX10-NEXT: early-clobber %7:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
 647     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %7.sub4_sub5_sub6_sub7
 648     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %7.sub0_sub1_sub2_sub3
 649     ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
 650     ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
 651     ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY [[COPY2]].sub2_sub3
 652     ; GFX10-NEXT: [[COPY6:%[0-9]+]]:sgpr_64 = COPY killed [[COPY2]].sub0_sub1
 653     ; GFX10-NEXT: S_ENDPGM 0
 654     ;
 655     ; GFX12-LABEL: name: merge_s_buffer_load_x8_out_of_x2ec_x2
 656     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 657     ; GFX12-NEXT: {{  $}}
 658     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 659     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 8)
 660     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 661     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 662     ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY [[COPY1]].sub0_sub1
 663     ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY killed [[COPY1]].sub2_sub3
 664     ; GFX12-NEXT: [[COPY5:%[0-9]+]]:sgpr_64 = COPY [[COPY2]].sub2_sub3
 665     ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sgpr_64 = COPY killed [[COPY2]].sub0_sub1
 666     ; GFX12-NEXT: S_ENDPGM 0
 667     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 668     early-clobber %1:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s64))
 669     early-clobber %2:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s64))
 670     %3:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s64))
 671     %4:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s64))
 672
 673     S_ENDPGM 0
 674 ...
 675 ---
 676
 677 name: merge_s_buffer_load_x8_out_of_x4ec
 678 tracksRegLiveness: true
 679 body:             |
 680   bb.0:
 681     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 682
 683     ; GFX10-LABEL: name: merge_s_buffer_load_x8_out_of_x4ec
 684     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 685     ; GFX10-NEXT: {{  $}}
 686     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 687     ; GFX10-NEXT: early-clobber %3:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 688     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %3.sub0_sub1_sub2_sub3
 689     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %3.sub4_sub5_sub6_sub7
 690     ; GFX10-NEXT: S_ENDPGM 0
 691     ;
 692     ; GFX12-LABEL: name: merge_s_buffer_load_x8_out_of_x4ec
 693     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 694     ; GFX12-NEXT: {{  $}}
 695     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 696     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 697     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 698     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 699     ; GFX12-NEXT: S_ENDPGM 0
 700     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 701     early-clobber %1:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM_ec %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s128))
 702     early-clobber %2:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM_ec %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s128))
 703
 704     S_ENDPGM 0
 705 ...
 706 ---
 707
 708 name: merge_s_buffer_load_x8_out_of_x4ec_x4
 709 tracksRegLiveness: true
 710 body:             |
 711   bb.0:
 712     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 713
 714     ; GFX10-LABEL: name: merge_s_buffer_load_x8_out_of_x4ec_x4
 715     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 716     ; GFX10-NEXT: {{  $}}
 717     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 718     ; GFX10-NEXT: early-clobber %3:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 719     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %3.sub0_sub1_sub2_sub3
 720     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %3.sub4_sub5_sub6_sub7
 721     ; GFX10-NEXT: S_ENDPGM 0
 722     ;
 723     ; GFX12-LABEL: name: merge_s_buffer_load_x8_out_of_x4ec_x4
 724     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 725     ; GFX12-NEXT: {{  $}}
 726     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 727     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 728     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 729     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 730     ; GFX12-NEXT: S_ENDPGM 0
 731     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 732     early-clobber %1:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM_ec %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s128))
 733     %2:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s128))
 734
 735     S_ENDPGM 0
 736 ...
 737 ---
 738
 739 name: merge_s_buffer_load_x8_out_of_x4_x4ec
 740 tracksRegLiveness: true
 741 body:             |
 742   bb.0:
 743     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 744
 745     ; GFX10-LABEL: name: merge_s_buffer_load_x8_out_of_x4_x4ec
 746     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 747     ; GFX10-NEXT: {{  $}}
 748     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 749     ; GFX10-NEXT: early-clobber %3:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 750     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %3.sub0_sub1_sub2_sub3
 751     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %3.sub4_sub5_sub6_sub7
 752     ; GFX10-NEXT: S_ENDPGM 0
 753     ;
 754     ; GFX12-LABEL: name: merge_s_buffer_load_x8_out_of_x4_x4ec
 755     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 756     ; GFX12-NEXT: {{  $}}
 757     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 758     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 759     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 760     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 761     ; GFX12-NEXT: S_ENDPGM 0
 762     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 763     %1:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s128))
 764     early-clobber %2:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM_ec %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s128))
 765
 766     S_ENDPGM 0
 767 ...
 768 ---
 769
 770 name: merge_s_buffer_load_x8_mixed_including_ec_opcodes
 771 tracksRegLiveness: true
 772 body:             |
 773   bb.0:
 774     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 775
 776     ; GFX10-LABEL: name: merge_s_buffer_load_x8_mixed_including_ec_opcodes
 777     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 778     ; GFX10-NEXT: {{  $}}
 779     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 780     ; GFX10-NEXT: early-clobber %7:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM_ec [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 781     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY %7.sub0_sub1_sub2_sub3
 782     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed %7.sub4_sub5_sub6_sub7
 783     ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY [[COPY2]].sub0_sub1
 784     ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY killed [[COPY2]].sub2_sub3
 785     ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY3]].sub0
 786     ; GFX10-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub1
 787     ; GFX10-NEXT: S_ENDPGM 0
 788     ;
 789     ; GFX12-LABEL: name: merge_s_buffer_load_x8_mixed_including_ec_opcodes
 790     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 791     ; GFX12-NEXT: {{  $}}
 792     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 793     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256), align 16)
 794     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 795     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 796     ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY [[COPY2]].sub0_sub1
 797     ; GFX12-NEXT: [[COPY4:%[0-9]+]]:sgpr_64 = COPY killed [[COPY2]].sub2_sub3
 798     ; GFX12-NEXT: [[COPY5:%[0-9]+]]:sreg_32_xm0_xexec = COPY [[COPY3]].sub0
 799     ; GFX12-NEXT: [[COPY6:%[0-9]+]]:sreg_32_xm0_xexec = COPY killed [[COPY3]].sub1
 800     ; GFX12-NEXT: S_ENDPGM 0
 801     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 802     early-clobber %1:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM_ec %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s128))
 803     %2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s32))
 804     early-clobber %3:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM_ec %0:sgpr_128, 24, 0 :: (dereferenceable invariant load (s64))
 805     %4:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0:sgpr_128, 20, 0 :: (dereferenceable invariant load (s32))
 806
 807     S_ENDPGM 0
 808 ...
 809 ---
 810
 811 name: merge_s_buffer_load_sgpr_imm_x2ec_x2ec
 812 tracksRegLiveness: true
 813 body:             |
 814   bb.0:
 815     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
 816
 817     ; GFX10-LABEL: name: merge_s_buffer_load_sgpr_imm_x2ec_x2ec
 818     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
 819     ; GFX10-NEXT: {{  $}}
 820     ; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 821     ; GFX10-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr4
 822     ; GFX10-NEXT: early-clobber %4:sgpr_128 = S_BUFFER_LOAD_DWORDX4_SGPR_IMM_ec [[COPY]], [[COPY1]], 0, 0 :: (dereferenceable invariant load (s128), align 8)
 823     ; GFX10-NEXT: [[COPY2:%[0-9]+]]:sgpr_64 = COPY %4.sub0_sub1
 824     ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY killed %4.sub2_sub3
 825     ; GFX10-NEXT: S_ENDPGM 0
 826     ;
 827     ; GFX12-LABEL: name: merge_s_buffer_load_sgpr_imm_x2ec_x2ec
 828     ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
 829     ; GFX12-NEXT: {{  $}}
 830     ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 831     ; GFX12-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr4
 832     ; GFX12-NEXT: [[S_BUFFER_LOAD_DWORDX4_SGPR_IMM:%[0-9]+]]:sgpr_128 = S_BUFFER_LOAD_DWORDX4_SGPR_IMM [[COPY]], [[COPY1]], 0, 0 :: (dereferenceable invariant load (s128), align 8)
 833     ; GFX12-NEXT: [[COPY2:%[0-9]+]]:sgpr_64 = COPY [[S_BUFFER_LOAD_DWORDX4_SGPR_IMM]].sub0_sub1
 834     ; GFX12-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY killed [[S_BUFFER_LOAD_DWORDX4_SGPR_IMM]].sub2_sub3
 835     ; GFX12-NEXT: S_ENDPGM 0
 836     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 837     %1:sreg_32 = COPY $sgpr4
 838     early-clobber %2:sgpr_64 = S_BUFFER_LOAD_DWORDX2_SGPR_IMM_ec %0:sgpr_128, %1:sreg_32, 0, 0 :: (dereferenceable invariant load (s64))
 839     early-clobber %3:sgpr_64 = S_BUFFER_LOAD_DWORDX2_SGPR_IMM_ec %0:sgpr_128, %1:sreg_32, 8, 0 :: (dereferenceable invariant load (s64))
 840
 841     S_ENDPGM 0
 842 ...
 843
 844 # No constrained opcode required when the MEM operand has met the required alignment.
 845
 846 ---
 847
 848 name: merge_s_buffer_load_x2_x2_no_constrained_opc_needed
 849 tracksRegLiveness: true
 850 body:             |
 851   bb.0:
 852     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 853
 854     ; CHECK-LABEL: name: merge_s_buffer_load_x2_x2_no_constrained_opc_needed
 855     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 856     ; CHECK-NEXT: {{  $}}
 857     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 858     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s128))
 859     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr_64 = COPY [[S_BUFFER_LOAD_DWORDX4_IMM]].sub0_sub1
 860     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr_64 = COPY killed [[S_BUFFER_LOAD_DWORDX4_IMM]].sub2_sub3
 861     ; CHECK-NEXT: S_ENDPGM 0
 862     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 863     %1:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s64), align 16)
 864     %2:sgpr_64 = S_BUFFER_LOAD_DWORDX2_IMM %0:sgpr_128, 8, 0 :: (dereferenceable invariant load (s64))
 865
 866     S_ENDPGM 0
 867 ...
 868 ---
 869
 870 name: merge_s_buffer_load_x4_x4_no_constrained_opc_needed
 871 tracksRegLiveness: true
 872 body:             |
 873   bb.0:
 874     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 875
 876     ; CHECK-LABEL: name: merge_s_buffer_load_x4_x4_no_constrained_opc_needed
 877     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 878     ; CHECK-NEXT: {{  $}}
 879     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 880     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_BUFFER_LOAD_DWORDX8_IMM [[COPY]], 0, 0 :: (dereferenceable invariant load (s256))
 881     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr_128 = COPY [[S_BUFFER_LOAD_DWORDX8_IMM]].sub0_sub1_sub2_sub3
 882     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr_128 = COPY killed [[S_BUFFER_LOAD_DWORDX8_IMM]].sub4_sub5_sub6_sub7
 883     ; CHECK-NEXT: S_ENDPGM 0
 884     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 885     %1:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM %0:sgpr_128, 0, 0 :: (dereferenceable invariant load (s128), align 32)
 886     %2:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM %0:sgpr_128, 16, 0 :: (dereferenceable invariant load (s128))
 887
 888     S_ENDPGM 0
 889 ...
 890 ---
 891
 892 name: merge_s_buffer_load_sgpr_imm_x2ec_x2ec_no_constrained_opc_needed
 893 tracksRegLiveness: true
 894 body:             |
 895   bb.0:
 896     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
 897
 898     ; CHECK-LABEL: name: merge_s_buffer_load_sgpr_imm_x2ec_x2ec_no_constrained_opc_needed
 899     ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
 900     ; CHECK-NEXT: {{  $}}
 901     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 902     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr4
 903     ; CHECK-NEXT: [[S_BUFFER_LOAD_DWORDX4_SGPR_IMM:%[0-9]+]]:sgpr_128 = S_BUFFER_LOAD_DWORDX4_SGPR_IMM [[COPY]], [[COPY1]], 0, 0 :: (dereferenceable invariant load (s128))
 904     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr_64 = COPY [[S_BUFFER_LOAD_DWORDX4_SGPR_IMM]].sub0_sub1
 905     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sgpr_64 = COPY killed [[S_BUFFER_LOAD_DWORDX4_SGPR_IMM]].sub2_sub3
 906     ; CHECK-NEXT: S_ENDPGM 0
 907     %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 908     %1:sreg_32 = COPY $sgpr4
 909     %2:sgpr_64 = S_BUFFER_LOAD_DWORDX2_SGPR_IMM %0:sgpr_128, %1:sreg_32, 0, 0 :: (dereferenceable invariant load (s64), align 16)
 910     %3:sgpr_64 = S_BUFFER_LOAD_DWORDX2_SGPR_IMM %0:sgpr_128, %1:sreg_32, 8, 0 :: (dereferenceable invariant load (s64))
 911
 912     S_ENDPGM 0
 913 ...