llvm/test/CodeGen/AMDGPU/si-fold-aligned-agprs.mir

   1 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
   2 # RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -run-pass=si-fold-operands -verify-machineinstrs -o - %s | FileCheck --check-prefix=GFX90A %s
   3
   4 # Should fold only the aligned agpr copies with their uses.
   5
   6 ---
   7 name:            aligned_agpr_64
   8 tracksRegLiveness: true
   9 body:             |
  10   bb.0:
  11     liveins: $vgpr0_vgpr1
  12     ; GFX90A-LABEL: name: aligned_agpr_64
  13     ; GFX90A: liveins: $vgpr0_vgpr1
  14     ; GFX90A-NEXT: {{  $}}
  15     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
  16     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_64_align2 = IMPLICIT_DEF
  17     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]], 0, 0, implicit $exec
  18     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
  19     %1:areg_64_align2 = IMPLICIT_DEF
  20     %2:areg_64_align2 = COPY killed %1
  21     GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
  22 ...
  23
  24 ---
  25 name:            unaligned_agpr_64
  26 tracksRegLiveness: true
  27 body:             |
  28   bb.0:
  29     liveins: $vgpr0_vgpr1
  30     ; GFX90A-LABEL: name: unaligned_agpr_64
  31     ; GFX90A: liveins: $vgpr0_vgpr1
  32     ; GFX90A-NEXT: {{  $}}
  33     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
  34     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_64 = IMPLICIT_DEF
  35     ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_64_align2 = COPY killed [[DEF]]
  36     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
  37     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
  38     %1:areg_64 = IMPLICIT_DEF
  39     %2:areg_64_align2 = COPY killed %1
  40     GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
  41 ...
  42
  43 ---
  44 name:            aligned_agpr_96_sub0_subg1
  45 tracksRegLiveness: true
  46 body:             |
  47   bb.0:
  48     liveins: $vgpr0_vgpr1
  49     ; GFX90A-LABEL: name: aligned_agpr_96_sub0_subg1
  50     ; GFX90A: liveins: $vgpr0_vgpr1
  51     ; GFX90A-NEXT: {{  $}}
  52     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
  53     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96_align2 = IMPLICIT_DEF
  54     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub0_sub1, 0, 0, implicit $exec
  55     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
  56     %1:areg_96_align2 = IMPLICIT_DEF
  57     %2:areg_64_align2 = COPY killed %1.sub0_sub1:areg_96_align2
  58     GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
  59 ...
  60
  61 ---
  62 name:            unaligned_agpr_96_sub1_sub2
  63 tracksRegLiveness: true
  64 body:             |
  65   bb.0:
  66     liveins: $vgpr0_vgpr1
  67     ; GFX90A-LABEL: name: unaligned_agpr_96_sub1_sub2
  68     ; GFX90A: liveins: $vgpr0_vgpr1
  69     ; GFX90A-NEXT: {{  $}}
  70     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
  71     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96_align2 = IMPLICIT_DEF
  72     ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_64_align2 = COPY killed [[DEF]].sub1_sub2
  73     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
  74     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
  75     %1:areg_96_align2 = IMPLICIT_DEF
  76     %2:areg_64_align2 = COPY killed %1.sub1_sub2:areg_96_align2
  77     GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
  78 ...
  79
  80 ---
  81 name:            aligned_agpr_96
  82 tracksRegLiveness: true
  83 body:             |
  84   bb.0:
  85     liveins: $vgpr0_vgpr1
  86     ; GFX90A-LABEL: name: aligned_agpr_96
  87     ; GFX90A: liveins: $vgpr0_vgpr1
  88     ; GFX90A-NEXT: {{  $}}
  89     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
  90     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96_align2 = IMPLICIT_DEF
  91     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]], 0, 0, implicit $exec
  92     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
  93     %1:areg_96_align2 = IMPLICIT_DEF
  94     %2:areg_96_align2 = COPY killed %1
  95     GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
  96 ...
  97
  98 ---
  99 name:            unaligned_agpr_96
 100 tracksRegLiveness: true
 101 body:             |
 102   bb.0:
 103     liveins: $vgpr0_vgpr1
 104     ; GFX90A-LABEL: name: unaligned_agpr_96
 105     ; GFX90A: liveins: $vgpr0_vgpr1
 106     ; GFX90A-NEXT: {{  $}}
 107     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
 108     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96 = IMPLICIT_DEF
 109     ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_96_align2 = COPY killed [[DEF]]
 110     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[COPY1]], 0, 0, implicit $exec
 111     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
 112     %1:areg_96 = IMPLICIT_DEF
 113     %2:areg_96_align2 = COPY killed %1
 114     GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
 115 ...
 116
 117 ---
 118 name:            aligned_agpr_128_sub0_sub1
 119 tracksRegLiveness: true
 120 body:             |
 121   bb.0:
 122     liveins: $vgpr0_vgpr1
 123     ; GFX90A-LABEL: name: aligned_agpr_128_sub0_sub1
 124     ; GFX90A: liveins: $vgpr0_vgpr1
 125     ; GFX90A-NEXT: {{  $}}
 126     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
 127     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
 128     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub0_sub1, 0, 0, implicit $exec
 129     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
 130     %1:areg_128_align2 = IMPLICIT_DEF
 131     %2:areg_64_align2 = COPY killed %1.sub0_sub1:areg_128_align2
 132     GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
 133 ...
 134
 135 ---
 136 name:            aligned_agpr_128_sub2_sub3
 137 tracksRegLiveness: true
 138 body:             |
 139   bb.0:
 140     liveins: $vgpr0_vgpr1
 141     ; GFX90A-LABEL: name: aligned_agpr_128_sub2_sub3
 142     ; GFX90A: liveins: $vgpr0_vgpr1
 143     ; GFX90A-NEXT: {{  $}}
 144     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
 145     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
 146     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub2_sub3, 0, 0, implicit $exec
 147     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
 148     %1:areg_128_align2 = IMPLICIT_DEF
 149     %2:areg_64_align2 = COPY killed %1.sub2_sub3:areg_128_align2
 150     GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
 151 ...
 152
 153 ---
 154 name:            unaligned_agpr_128_sub1_sub2
 155 tracksRegLiveness: true
 156 body:             |
 157   bb.0:
 158     liveins: $vgpr0_vgpr1
 159     ; GFX90A-LABEL: name: unaligned_agpr_128_sub1_sub2
 160     ; GFX90A: liveins: $vgpr0_vgpr1
 161     ; GFX90A-NEXT: {{  $}}
 162     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
 163     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
 164     ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_64_align2 = COPY killed [[DEF]].sub1_sub2
 165     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
 166     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
 167     %1:areg_128_align2 = IMPLICIT_DEF
 168     %2:areg_64_align2 = COPY killed %1.sub1_sub2:areg_128_align2
 169     GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
 170 ...
 171
 172 ---
 173 name:            aligned_agpr_128_sub0_sub1_sub2
 174 tracksRegLiveness: true
 175 body:             |
 176   bb.0:
 177     liveins: $vgpr0_vgpr1
 178     ; GFX90A-LABEL: name: aligned_agpr_128_sub0_sub1_sub2
 179     ; GFX90A: liveins: $vgpr0_vgpr1
 180     ; GFX90A-NEXT: {{  $}}
 181     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
 182     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
 183     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]].sub0_sub1_sub2, 0, 0, implicit $exec
 184     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
 185     %1:areg_128_align2 = IMPLICIT_DEF
 186     %2:areg_96_align2 = COPY killed %1.sub0_sub1_sub2:areg_128_align2
 187     GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
 188 ...
 189
 190 ---
 191 name:            unaligned_agpr_128_sub1_sub2_sub3
 192 tracksRegLiveness: true
 193 body:             |
 194   bb.0:
 195     liveins: $vgpr0_vgpr1
 196     ; GFX90A-LABEL: name: unaligned_agpr_128_sub1_sub2_sub3
 197     ; GFX90A: liveins: $vgpr0_vgpr1
 198     ; GFX90A-NEXT: {{  $}}
 199     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
 200     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
 201     ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_96_align2 = COPY killed [[DEF]].sub1_sub2_sub3
 202     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[COPY1]], 0, 0, implicit $exec
 203     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
 204     %1:areg_128_align2 = IMPLICIT_DEF
 205     %2:areg_96_align2 = COPY killed %1.sub1_sub2_sub3:areg_128_align2
 206     GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
 207 ...
 208
 209 ---
 210 name:            aligned_agpr_128
 211 tracksRegLiveness: true
 212 body:             |
 213   bb.0:
 214     liveins: $vgpr0_vgpr1
 215     ; GFX90A-LABEL: name: aligned_agpr_128
 216     ; GFX90A: liveins: $vgpr0_vgpr1
 217     ; GFX90A-NEXT: {{  $}}
 218     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
 219     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
 220     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX4 [[COPY]], [[DEF]], 0, 0, implicit $exec
 221     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
 222     %1:areg_128_align2 = IMPLICIT_DEF
 223     %2:areg_128_align2 = COPY killed %1
 224     GLOBAL_STORE_DWORDX4 %0, %2, 0, 0, implicit $exec
 225 ...
 226
 227 ---
 228 name:            unaligned_agpr_128
 229 tracksRegLiveness: true
 230 body:             |
 231   bb.0:
 232     liveins: $vgpr0_vgpr1
 233     ; GFX90A-LABEL: name: unaligned_agpr_128
 234     ; GFX90A: liveins: $vgpr0_vgpr1
 235     ; GFX90A-NEXT: {{  $}}
 236     ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
 237     ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128 = IMPLICIT_DEF
 238     ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_128_align2 = COPY killed [[DEF]]
 239     ; GFX90A-NEXT: GLOBAL_STORE_DWORDX4 [[COPY]], [[COPY1]], 0, 0, implicit $exec
 240     %0:vreg_64_align2 = COPY $vgpr0_vgpr1
 241     %1:areg_128 = IMPLICIT_DEF
 242     %2:areg_128_align2 = COPY killed %1
 243     GLOBAL_STORE_DWORDX4 %0, %2, 0, 0, implicit $exec
 244 ...