llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn.mfma.mir

   1 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
   2 # RUN: llc -mtriple=amdgcn -mcpu=gfx908 -run-pass=amdgpu-regbankselect -regbankselect-fast -verify-machineinstrs %s -o - | FileCheck %s
   3 # RUN: llc -mtriple=amdgcn -mcpu=gfx908 -run-pass=amdgpu-regbankselect -regbankselect-greedy -verify-machineinstrs %s -o - | FileCheck %s
   4
   5 ---
   6 name: mfma_f32_32x32x1f32_vva
   7 legalized: true
   8 tracksRegLiveness: true
   9 body: |
  10   bb.0:
  11     liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
  12
  13     ; CHECK-LABEL: name: mfma_f32_32x32x1f32_vva
  14     ; CHECK: liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
  15     ; CHECK-NEXT: {{  $}}
  16     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
  17     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
  18     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<32 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
  19     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x1f32), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<32 x s32>), 0, 0, 0
  20     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY [[INT]](<32 x s32>)
  21     %0:_(s32) = COPY $vgpr0
  22     %1:_(s32) = COPY $vgpr1
  23     %2:_(<32 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
  24     %3:_(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x1f32), %0, %1, %2, 0, 0, 0
  25     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY %3
  26 ...
  27
  28 ---
  29 name: mfma_f32_32x32x1f32_sss
  30 legalized: true
  31 tracksRegLiveness: true
  32 body: |
  33   bb.0:
  34     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
  35
  36     ; CHECK-LABEL: name: mfma_f32_32x32x1f32_sss
  37     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
  38     ; CHECK-NEXT: {{  $}}
  39     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
  40     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
  41     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<32 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
  42     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
  43     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
  44     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<32 x s32>) = COPY [[COPY2]](<32 x s32>)
  45     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x1f32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<32 x s32>), 0, 0, 0
  46     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY [[INT]](<32 x s32>)
  47     %0:_(s32) = COPY $sgpr32
  48     %1:_(s32) = COPY $sgpr33
  49     %2:_(<32 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
  50     %3:_(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x1f32), %0, %1, %2, 0, 0, 0
  51     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY %3
  52 ...
  53
  54 ---
  55 name: mfma_f32_16x16x1f32_vva
  56 legalized: true
  57 tracksRegLiveness: true
  58 body: |
  59   bb.0:
  60     liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
  61
  62     ; CHECK-LABEL: name: mfma_f32_16x16x1f32_vva
  63     ; CHECK: liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
  64     ; CHECK-NEXT: {{  $}}
  65     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
  66     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
  67     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
  68     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x1f32), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<16 x s32>), 0, 0, 0
  69     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
  70     %0:_(s32) = COPY $vgpr0
  71     %1:_(s32) = COPY $vgpr1
  72     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
  73     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x1f32), %0, %1, %2, 0, 0, 0
  74     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
  75 ...
  76
  77 ---
  78 name: mfma_f32_16x16x1f32_sss
  79 legalized: true
  80 tracksRegLiveness: true
  81 body: |
  82   bb.0:
  83     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
  84
  85     ; CHECK-LABEL: name: mfma_f32_16x16x1f32_sss
  86     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
  87     ; CHECK-NEXT: {{  $}}
  88     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
  89     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
  90     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
  91     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
  92     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
  93     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
  94     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x1f32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<16 x s32>), 0, 0, 0
  95     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
  96     %0:_(s32) = COPY $sgpr32
  97     %1:_(s32) = COPY $sgpr33
  98     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
  99     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x1f32), %0, %1, %2, 0, 0, 0
 100     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 101 ...
 102
 103 ---
 104 name: mfma_f32_4x4x1f32_vva
 105 legalized: true
 106 tracksRegLiveness: true
 107 body: |
 108   bb.0:
 109     liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3
 110
 111     ; CHECK-LABEL: name: mfma_f32_4x4x1f32_vva
 112     ; CHECK: liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3
 113     ; CHECK-NEXT: {{  $}}
 114     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 115     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
 116     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 117     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x1f32), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<4 x s32>), 0, 0, 0
 118     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 119     %0:_(s32) = COPY $vgpr0
 120     %1:_(s32) = COPY $vgpr1
 121     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 122     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x1f32), %0, %1, %2, 0, 0, 0
 123     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 124 ...
 125
 126 ---
 127 name: mfma_f32_4x4x1f32_sss
 128 legalized: true
 129 tracksRegLiveness: true
 130 body: |
 131   bb.0:
 132     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 133
 134     ; CHECK-LABEL: name: mfma_f32_4x4x1f32_sss
 135     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 136     ; CHECK-NEXT: {{  $}}
 137     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 138     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 139     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 140     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 141     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 142     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 143     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x1f32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<4 x s32>), 0, 0, 0
 144     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 145     %0:_(s32) = COPY $sgpr32
 146     %1:_(s32) = COPY $sgpr33
 147     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 148     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x1f32), %0, %1, %2, 0, 0, 0
 149     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 150 ...
 151
 152 ---
 153 name: mfma_f32_32x32x2f32_vva
 154 legalized: true
 155 tracksRegLiveness: true
 156 body: |
 157   bb.0:
 158     liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 159
 160     ; CHECK-LABEL: name: mfma_f32_32x32x2f32_vva
 161     ; CHECK: liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 162     ; CHECK-NEXT: {{  $}}
 163     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 164     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
 165     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 166     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x2f32), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<16 x s32>), 0, 0, 0
 167     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 168     %0:_(s32) = COPY $vgpr0
 169     %1:_(s32) = COPY $vgpr1
 170     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 171     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x2f32), %0, %1, %2, 0, 0, 0
 172     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 173 ...
 174
 175 ---
 176 name: mfma_f32_32x32x2f32_sss
 177 legalized: true
 178 tracksRegLiveness: true
 179 body: |
 180   bb.0:
 181     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 182
 183     ; CHECK-LABEL: name: mfma_f32_32x32x2f32_sss
 184     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 185     ; CHECK-NEXT: {{  $}}
 186     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 187     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 188     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 189     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 190     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 191     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 192     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x2f32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<16 x s32>), 0, 0, 0
 193     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 194     %0:_(s32) = COPY $sgpr32
 195     %1:_(s32) = COPY $sgpr33
 196     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 197     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x2f32), %0, %1, %2, 0, 0, 0
 198     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 199 ...
 200
 201 ---
 202 name: mfma_f32_16x16x4f32_vva
 203 legalized: true
 204 tracksRegLiveness: true
 205 body: |
 206   bb.0:
 207     liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3
 208
 209     ; CHECK-LABEL: name: mfma_f32_16x16x4f32_vva
 210     ; CHECK: liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3
 211     ; CHECK-NEXT: {{  $}}
 212     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 213     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
 214     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 215     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f32), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<4 x s32>), 0, 0, 0
 216     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 217     %0:_(s32) = COPY $vgpr0
 218     %1:_(s32) = COPY $vgpr1
 219     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 220     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f32), %0, %1, %2, 0, 0, 0
 221     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 222 ...
 223
 224 ---
 225 name: mfma_f32_16x16x4f32_sss
 226 legalized: true
 227 tracksRegLiveness: true
 228 body: |
 229   bb.0:
 230     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 231
 232     ; CHECK-LABEL: name: mfma_f32_16x16x4f32_sss
 233     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 234     ; CHECK-NEXT: {{  $}}
 235     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 236     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 237     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 238     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 239     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 240     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 241     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<4 x s32>), 0, 0, 0
 242     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 243     %0:_(s32) = COPY $sgpr32
 244     %1:_(s32) = COPY $sgpr33
 245     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 246     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f32), %0, %1, %2, 0, 0, 0
 247     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 248 ...
 249
 250 ---
 251 name: mfma_f32_32x32x4f16_vva
 252 legalized: true
 253 tracksRegLiveness: true
 254 body: |
 255   bb.0:
 256     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 257
 258     ; CHECK-LABEL: name: mfma_f32_32x32x4f16_vva
 259     ; CHECK: liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 260     ; CHECK-NEXT: {{  $}}
 261     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
 262     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
 263     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<32 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 264     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x4f16), [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<32 x s32>), 0, 0, 0
 265     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY [[INT]](<32 x s32>)
 266     %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
 267     %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
 268     %2:_(<32 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 269     %3:_(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x4f16), %0, %1, %2, 0, 0, 0
 270     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY %3
 271 ...
 272
 273 ---
 274 name: mfma_f32_32x32x4f16_sss
 275 legalized: true
 276 tracksRegLiveness: true
 277 body: |
 278   bb.0:
 279     liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 280
 281     ; CHECK-LABEL: name: mfma_f32_32x32x4f16_sss
 282     ; CHECK: liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 283     ; CHECK-NEXT: {{  $}}
 284     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr32_sgpr33
 285     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr34_sgpr35
 286     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<32 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 287     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY]](<4 x s16>)
 288     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY1]](<4 x s16>)
 289     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<32 x s32>) = COPY [[COPY2]](<32 x s32>)
 290     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x4f16), [[COPY3]](<4 x s16>), [[COPY4]](<4 x s16>), [[COPY5]](<32 x s32>), 0, 0, 0
 291     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY [[INT]](<32 x s32>)
 292     %0:_(<4 x s16>) = COPY $sgpr32_sgpr33
 293     %1:_(<4 x s16>) = COPY $sgpr34_sgpr35
 294     %2:_(<32 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 295     %3:_(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x4f16), %0, %1, %2, 0, 0, 0
 296     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY %3
 297 ...
 298
 299 ---
 300 name: mfma_f32_16x16x4f16_vva
 301 legalized: true
 302 tracksRegLiveness: true
 303 body: |
 304   bb.0:
 305     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 306
 307     ; CHECK-LABEL: name: mfma_f32_16x16x4f16_vva
 308     ; CHECK: liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 309     ; CHECK-NEXT: {{  $}}
 310     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
 311     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
 312     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 313     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f16), [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<16 x s32>), 0, 0, 0
 314     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 315     %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
 316     %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
 317     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 318     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f16), %0, %1, %2, 0, 0, 0
 319     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 320 ...
 321
 322 ---
 323 name: mfma_f32_16x16x4f16_sss
 324 legalized: true
 325 tracksRegLiveness: true
 326 body: |
 327   bb.0:
 328     liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 329
 330     ; CHECK-LABEL: name: mfma_f32_16x16x4f16_sss
 331     ; CHECK: liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 332     ; CHECK-NEXT: {{  $}}
 333     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr32_sgpr33
 334     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr34_sgpr35
 335     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 336     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY]](<4 x s16>)
 337     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY1]](<4 x s16>)
 338     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 339     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f16), [[COPY3]](<4 x s16>), [[COPY4]](<4 x s16>), [[COPY5]](<16 x s32>), 0, 0, 0
 340     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 341     %0:_(<4 x s16>) = COPY $sgpr32_sgpr33
 342     %1:_(<4 x s16>) = COPY $sgpr34_sgpr35
 343     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 344     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f16), %0, %1, %2, 0, 0, 0
 345     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 346 ...
 347
 348 ---
 349 name: mfma_f32_4x4x4f16_vva
 350 legalized: true
 351 tracksRegLiveness: true
 352 body: |
 353   bb.0:
 354     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3
 355
 356     ; CHECK-LABEL: name: mfma_f32_4x4x4f16_vva
 357     ; CHECK: liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3
 358     ; CHECK-NEXT: {{  $}}
 359     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
 360     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
 361     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 362     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x4f16), [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<4 x s32>), 0, 0, 0
 363     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 364     %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
 365     %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
 366     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 367     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x4f16), %0, %1, %2, 0, 0, 0
 368     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 369 ...
 370
 371 ---
 372 name: mfma_f32_4x4x4f16_sss
 373 legalized: true
 374 tracksRegLiveness: true
 375 body: |
 376   bb.0:
 377     liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3
 378
 379     ; CHECK-LABEL: name: mfma_f32_4x4x4f16_sss
 380     ; CHECK: liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3
 381     ; CHECK-NEXT: {{  $}}
 382     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr32_sgpr33
 383     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr34_sgpr35
 384     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 385     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY]](<4 x s16>)
 386     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY1]](<4 x s16>)
 387     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 388     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x4f16), [[COPY3]](<4 x s16>), [[COPY4]](<4 x s16>), [[COPY5]](<4 x s32>), 0, 0, 0
 389     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 390     %0:_(<4 x s16>) = COPY $sgpr32_sgpr33
 391     %1:_(<4 x s16>) = COPY $sgpr34_sgpr35
 392     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 393     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x4f16), %0, %1, %2, 0, 0, 0
 394     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 395 ...
 396
 397 ---
 398 name: mfma_f32_32x32x8f16_vva
 399 legalized: true
 400 tracksRegLiveness: true
 401 body: |
 402   bb.0:
 403     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 404
 405     ; CHECK-LABEL: name: mfma_f32_32x32x8f16_vva
 406     ; CHECK: liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 407     ; CHECK-NEXT: {{  $}}
 408     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
 409     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
 410     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 411     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x8f16), [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<16 x s32>), 0, 0, 0
 412     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 413     %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
 414     %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
 415     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 416     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x8f16), %0, %1, %2, 0, 0, 0
 417     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 418 ...
 419
 420 ---
 421 name: mfma_f32_32x32x8f16_sss
 422 legalized: true
 423 tracksRegLiveness: true
 424 body: |
 425   bb.0:
 426     liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 427
 428     ; CHECK-LABEL: name: mfma_f32_32x32x8f16_sss
 429     ; CHECK: liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 430     ; CHECK-NEXT: {{  $}}
 431     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr32_sgpr33
 432     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr34_sgpr35
 433     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 434     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY]](<4 x s16>)
 435     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY1]](<4 x s16>)
 436     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 437     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x8f16), [[COPY3]](<4 x s16>), [[COPY4]](<4 x s16>), [[COPY5]](<16 x s32>), 0, 0, 0
 438     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 439     %0:_(<4 x s16>) = COPY $sgpr32_sgpr33
 440     %1:_(<4 x s16>) = COPY $sgpr34_sgpr35
 441     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 442     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x8f16), %0, %1, %2, 0, 0, 0
 443     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 444 ...
 445
 446 ---
 447 name: mfma_f32_16x16x16f16_vva
 448 legalized: true
 449 tracksRegLiveness: true
 450 body: |
 451   bb.0:
 452     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3
 453
 454     ; CHECK-LABEL: name: mfma_f32_16x16x16f16_vva
 455     ; CHECK: liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3
 456     ; CHECK-NEXT: {{  $}}
 457     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
 458     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
 459     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 460     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x16f16), [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<4 x s32>), 0, 0, 0
 461     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 462     %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
 463     %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
 464     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 465     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x16f16), %0, %1, %2, 0, 0, 0
 466     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 467 ...
 468
 469 ---
 470 name: mfma_f32_16x16x16f16_sss
 471 legalized: true
 472 tracksRegLiveness: true
 473 body: |
 474   bb.0:
 475     liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3
 476
 477     ; CHECK-LABEL: name: mfma_f32_16x16x16f16_sss
 478     ; CHECK: liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3
 479     ; CHECK-NEXT: {{  $}}
 480     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr32_sgpr33
 481     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr34_sgpr35
 482     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 483     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY]](<4 x s16>)
 484     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY1]](<4 x s16>)
 485     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 486     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x16f16), [[COPY3]](<4 x s16>), [[COPY4]](<4 x s16>), [[COPY5]](<4 x s32>), 0, 0, 0
 487     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 488     %0:_(<4 x s16>) = COPY $sgpr32_sgpr33
 489     %1:_(<4 x s16>) = COPY $sgpr34_sgpr35
 490     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 491     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x16f16), %0, %1, %2, 0, 0, 0
 492     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 493 ...
 494
 495 ---
 496 name: mfma_i32_32x32x4i8_vva
 497 legalized: true
 498 tracksRegLiveness: true
 499 body: |
 500   bb.0:
 501     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 502
 503     ; CHECK-LABEL: name: mfma_i32_32x32x4i8_vva
 504     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 505     ; CHECK-NEXT: {{  $}}
 506     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 507     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
 508     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<32 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 509     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.32x32x4i8), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<32 x s32>), 0, 0, 0
 510     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY [[INT]](<32 x s32>)
 511     %0:_(s32) = COPY $vgpr0
 512     %1:_(s32) = COPY $vgpr2
 513     %2:_(<32 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 514     %3:_(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.32x32x4i8), %0, %1, %2, 0, 0, 0
 515     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY %3
 516 ...
 517
 518 ---
 519 name: mfma_i32_32x32x4i8_sss
 520 legalized: true
 521 tracksRegLiveness: true
 522 body: |
 523   bb.0:
 524     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 525
 526     ; CHECK-LABEL: name: mfma_i32_32x32x4i8_sss
 527     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 528     ; CHECK-NEXT: {{  $}}
 529     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 530     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 531     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<32 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 532     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 533     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 534     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<32 x s32>) = COPY [[COPY2]](<32 x s32>)
 535     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.32x32x4i8), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<32 x s32>), 0, 0, 0
 536     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY [[INT]](<32 x s32>)
 537     %0:_(s32) = COPY $sgpr32
 538     %1:_(s32) = COPY $sgpr33
 539     %2:_(<32 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 540     %3:_(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.32x32x4i8), %0, %1, %2, 0, 0, 0
 541     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY %3
 542 ...
 543
 544 ---
 545 name: mfma_i32_16x16x4i8_vva
 546 legalized: true
 547 tracksRegLiveness: true
 548 body: |
 549   bb.0:
 550     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 551
 552     ; CHECK-LABEL: name: mfma_i32_16x16x4i8_vva
 553     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 554     ; CHECK-NEXT: {{  $}}
 555     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 556     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
 557     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 558     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.16x16x4i8), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<16 x s32>), 0, 0, 0
 559     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 560     %0:_(s32) = COPY $vgpr0
 561     %1:_(s32) = COPY $vgpr2
 562     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 563     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.16x16x4i8), %0, %1, %2, 0, 0, 0
 564     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 565 ...
 566
 567 ---
 568 name: mfma_i32_16x16x4i8_sss
 569 legalized: true
 570 tracksRegLiveness: true
 571 body: |
 572   bb.0:
 573     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 574
 575     ; CHECK-LABEL: name: mfma_i32_16x16x4i8_sss
 576     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 577     ; CHECK-NEXT: {{  $}}
 578     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 579     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 580     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 581     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 582     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 583     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 584     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.16x16x4i8), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<16 x s32>), 0, 0, 0
 585     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 586     %0:_(s32) = COPY $sgpr32
 587     %1:_(s32) = COPY $sgpr33
 588     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 589     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.16x16x4i8), %0, %1, %2, 0, 0, 0
 590     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 591 ...
 592
 593 ---
 594 name: mfma_i32_4x4x4i8_vva
 595 legalized: true
 596 tracksRegLiveness: true
 597 body: |
 598   bb.0:
 599     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 600
 601     ; CHECK-LABEL: name: mfma_i32_4x4x4i8_vva
 602     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 603     ; CHECK-NEXT: {{  $}}
 604     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 605     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
 606     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 607     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.4x4x4i8), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<4 x s32>), 0, 0, 0
 608     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 609     %0:_(s32) = COPY $vgpr0
 610     %1:_(s32) = COPY $vgpr2
 611     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 612     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.4x4x4i8), %0, %1, %2, 0, 0, 0
 613     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 614 ...
 615
 616 ---
 617 name: mfma_i32_4x4x4i8_sss
 618 legalized: true
 619 tracksRegLiveness: true
 620 body: |
 621   bb.0:
 622     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 623
 624     ; CHECK-LABEL: name: mfma_i32_4x4x4i8_sss
 625     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 626     ; CHECK-NEXT: {{  $}}
 627     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 628     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 629     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 630     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 631     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 632     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 633     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.4x4x4i8), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<4 x s32>), 0, 0, 0
 634     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 635     %0:_(s32) = COPY $sgpr32
 636     %1:_(s32) = COPY $sgpr33
 637     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 638     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.4x4x4i8), %0, %1, %2, 0, 0, 0
 639     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 640 ...
 641
 642 ---
 643 name: mfma_i32_32x32x8i8_vva
 644 legalized: true
 645 tracksRegLiveness: true
 646 body: |
 647   bb.0:
 648     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 649
 650     ; CHECK-LABEL: name: mfma_i32_32x32x8i8_vva
 651     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 652     ; CHECK-NEXT: {{  $}}
 653     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 654     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
 655     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 656     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.32x32x8i8), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<16 x s32>), 0, 0, 0
 657     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 658     %0:_(s32) = COPY $vgpr0
 659     %1:_(s32) = COPY $vgpr2
 660     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 661     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.32x32x8i8), %0, %1, %2, 0, 0, 0
 662     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 663 ...
 664
 665 ---
 666 name: mfma_i32_32x32x8i8_sss
 667 legalized: true
 668 tracksRegLiveness: true
 669 body: |
 670   bb.0:
 671     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 672
 673     ; CHECK-LABEL: name: mfma_i32_32x32x8i8_sss
 674     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 675     ; CHECK-NEXT: {{  $}}
 676     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 677     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 678     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 679     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 680     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 681     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 682     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.32x32x8i8), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<16 x s32>), 0, 0, 0
 683     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 684     %0:_(s32) = COPY $sgpr32
 685     %1:_(s32) = COPY $sgpr33
 686     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 687     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.32x32x8i8), %0, %1, %2, 0, 0, 0
 688     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 689 ...
 690
 691 ---
 692 name: mfma_i32_16x16x16i8_vva
 693 legalized: true
 694 tracksRegLiveness: true
 695 body: |
 696   bb.0:
 697     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 698
 699     ; CHECK-LABEL: name: mfma_i32_16x16x16i8_vva
 700     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 701     ; CHECK-NEXT: {{  $}}
 702     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 703     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
 704     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 705     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.16x16x16i8), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<4 x s32>), 0, 0, 0
 706     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 707     %0:_(s32) = COPY $vgpr0
 708     %1:_(s32) = COPY $vgpr2
 709     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 710     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.16x16x16i8), %0, %1, %2, 0, 0, 0
 711     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 712 ...
 713
 714 ---
 715 name: mfma_i32_16x16x16i8_sss
 716 legalized: true
 717 tracksRegLiveness: true
 718 body: |
 719   bb.0:
 720     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 721
 722     ; CHECK-LABEL: name: mfma_i32_16x16x16i8_sss
 723     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 724     ; CHECK-NEXT: {{  $}}
 725     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 726     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 727     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 728     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 729     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 730     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 731     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.16x16x16i8), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<4 x s32>), 0, 0, 0
 732     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 733     %0:_(s32) = COPY $sgpr32
 734     %1:_(s32) = COPY $sgpr33
 735     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 736     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.i32.16x16x16i8), %0, %1, %2, 0, 0, 0
 737     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 738 ...
 739
 740 ---
 741 name: mfma_f32_32x32x2bf16_vva
 742 legalized: true
 743 tracksRegLiveness: true
 744 body: |
 745   bb.0:
 746     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 747
 748     ; CHECK-LABEL: name: mfma_f32_32x32x2bf16_vva
 749     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 750     ; CHECK-NEXT: {{  $}}
 751     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
 752     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr2
 753     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<32 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 754     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x2bf16), [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<32 x s32>), 0, 0, 0
 755     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY [[INT]](<32 x s32>)
 756     %0:_(<2 x s16>) = COPY $vgpr0
 757     %1:_(<2 x s16>) = COPY $vgpr2
 758     %2:_(<32 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
 759     %3:_(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x2bf16), %0, %1, %2, 0, 0, 0
 760     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY %3
 761 ...
 762
 763 ---
 764 name: mfma_f32_32x32x2bf16_sss
 765 legalized: true
 766 tracksRegLiveness: true
 767 body: |
 768   bb.0:
 769     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 770
 771     ; CHECK-LABEL: name: mfma_f32_32x32x2bf16_sss
 772     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 773     ; CHECK-NEXT: {{  $}}
 774     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr32
 775     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr33
 776     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<32 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 777     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY]](<2 x s16>)
 778     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY1]](<2 x s16>)
 779     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<32 x s32>) = COPY [[COPY2]](<32 x s32>)
 780     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x2bf16), [[COPY3]](<2 x s16>), [[COPY4]](<2 x s16>), [[COPY5]](<32 x s32>), 0, 0, 0
 781     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY [[INT]](<32 x s32>)
 782     %0:_(<2 x s16>) = COPY $sgpr32
 783     %1:_(<2 x s16>) = COPY $sgpr33
 784     %2:_(<32 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31
 785     %3:_(<32 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x2bf16), %0, %1, %2, 0, 0, 0
 786     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 = COPY %3
 787 ...
 788
 789 ---
 790 name: mfma_f32_16x16x2bf16_vva
 791 legalized: true
 792 tracksRegLiveness: true
 793 body: |
 794   bb.0:
 795     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 796
 797     ; CHECK-LABEL: name: mfma_f32_16x16x2bf16_vva
 798     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 799     ; CHECK-NEXT: {{  $}}
 800     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
 801     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr2
 802     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 803     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x2bf16), [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<16 x s32>), 0, 0, 0
 804     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 805     %0:_(<2 x s16>) = COPY $vgpr0
 806     %1:_(<2 x s16>) = COPY $vgpr2
 807     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 808     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x2bf16), %0, %1, %2, 0, 0, 0
 809     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 810 ...
 811
 812 ---
 813 name: mfma_f32_16x16x2bf16_sss
 814 legalized: true
 815 tracksRegLiveness: true
 816 body: |
 817   bb.0:
 818     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 819
 820     ; CHECK-LABEL: name: mfma_f32_16x16x2bf16_sss
 821     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 822     ; CHECK-NEXT: {{  $}}
 823     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr32
 824     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr33
 825     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 826     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY]](<2 x s16>)
 827     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY1]](<2 x s16>)
 828     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 829     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x2bf16), [[COPY3]](<2 x s16>), [[COPY4]](<2 x s16>), [[COPY5]](<16 x s32>), 0, 0, 0
 830     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 831     %0:_(<2 x s16>) = COPY $sgpr32
 832     %1:_(<2 x s16>) = COPY $sgpr33
 833     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 834     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x2bf16), %0, %1, %2, 0, 0, 0
 835     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 836 ...
 837
 838 ---
 839 name: mfma_f32_4x4x2bf16_vva
 840 legalized: true
 841 tracksRegLiveness: true
 842 body: |
 843   bb.0:
 844     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 845
 846     ; CHECK-LABEL: name: mfma_f32_4x4x2bf16_vva
 847     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 848     ; CHECK-NEXT: {{  $}}
 849     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
 850     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr2
 851     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 852     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x2bf16), [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<4 x s32>), 0, 0, 0
 853     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 854     %0:_(<2 x s16>) = COPY $vgpr0
 855     %1:_(<2 x s16>) = COPY $vgpr2
 856     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 857     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x2bf16), %0, %1, %2, 0, 0, 0
 858     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 859 ...
 860
 861 ---
 862 name: mfma_f32_4x4x2bf16_sss
 863 legalized: true
 864 tracksRegLiveness: true
 865 body: |
 866   bb.0:
 867     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 868
 869     ; CHECK-LABEL: name: mfma_f32_4x4x2bf16_sss
 870     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 871     ; CHECK-NEXT: {{  $}}
 872     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr32
 873     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr33
 874     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 875     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY]](<2 x s16>)
 876     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY1]](<2 x s16>)
 877     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 878     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x2bf16), [[COPY3]](<2 x s16>), [[COPY4]](<2 x s16>), [[COPY5]](<4 x s32>), 0, 0, 0
 879     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 880     %0:_(<2 x s16>) = COPY $sgpr32
 881     %1:_(<2 x s16>) = COPY $sgpr33
 882     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 883     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.4x4x2bf16), %0, %1, %2, 0, 0, 0
 884     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 885 ...
 886
 887 ---
 888 name: mfma_f32_32x32x4bf16_vva
 889 legalized: true
 890 tracksRegLiveness: true
 891 body: |
 892   bb.0:
 893     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 894
 895     ; CHECK-LABEL: name: mfma_f32_32x32x4bf16_vva
 896     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 897     ; CHECK-NEXT: {{  $}}
 898     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
 899     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr2
 900     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 901     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x4bf16), [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<16 x s32>), 0, 0, 0
 902     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 903     %0:_(<2 x s16>) = COPY $vgpr0
 904     %1:_(<2 x s16>) = COPY $vgpr2
 905     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 906     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x4bf16), %0, %1, %2, 0, 0, 0
 907     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 908 ...
 909
 910 ---
 911 name: mfma_f32_32x32x4bf16_sss
 912 legalized: true
 913 tracksRegLiveness: true
 914 body: |
 915   bb.0:
 916     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 917
 918     ; CHECK-LABEL: name: mfma_f32_32x32x4bf16_sss
 919     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 920     ; CHECK-NEXT: {{  $}}
 921     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr32
 922     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr33
 923     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 924     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY]](<2 x s16>)
 925     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY1]](<2 x s16>)
 926     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 927     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x4bf16), [[COPY3]](<2 x s16>), [[COPY4]](<2 x s16>), [[COPY5]](<16 x s32>), 0, 0, 0
 928     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 929     %0:_(<2 x s16>) = COPY $sgpr32
 930     %1:_(<2 x s16>) = COPY $sgpr33
 931     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 932     %3:_(<16 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.32x32x4bf16), %0, %1, %2, 0, 0, 0
 933     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 934 ...
 935
 936 ---
 937 name: mfma_f32_16x16x8bf16_vva
 938 legalized: true
 939 tracksRegLiveness: true
 940 body: |
 941   bb.0:
 942     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 943
 944     ; CHECK-LABEL: name: mfma_f32_16x16x8bf16_vva
 945     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 946     ; CHECK-NEXT: {{  $}}
 947     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
 948     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr2
 949     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 950     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x8bf16), [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<4 x s32>), 0, 0, 0
 951     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 952     %0:_(<2 x s16>) = COPY $vgpr0
 953     %1:_(<2 x s16>) = COPY $vgpr2
 954     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 955     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x8bf16), %0, %1, %2, 0, 0, 0
 956     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 957 ...
 958
 959 ---
 960 name: mfma_f32_16x16x8bf16_sss
 961 legalized: true
 962 tracksRegLiveness: true
 963 body: |
 964   bb.0:
 965     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 966
 967     ; CHECK-LABEL: name: mfma_f32_16x16x8bf16_sss
 968     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 969     ; CHECK-NEXT: {{  $}}
 970     ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr32
 971     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr33
 972     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 973     ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY]](<2 x s16>)
 974     ; CHECK-NEXT: [[COPY4:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY1]](<2 x s16>)
 975     ; CHECK-NEXT: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 976     ; CHECK-NEXT: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x8bf16), [[COPY3]](<2 x s16>), [[COPY4]](<2 x s16>), [[COPY5]](<4 x s32>), 0, 0, 0
 977     ; CHECK-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 978     %0:_(<2 x s16>) = COPY $sgpr32
 979     %1:_(<2 x s16>) = COPY $sgpr33
 980     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 981     %3:_(<4 x s32>) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.mfma.f32.16x16x8bf16), %0, %1, %2, 0, 0, 0
 982     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 983 ...