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 -march=amdgcn -mcpu=gfx908 -run-pass=regbankselect -regbankselect-fast -verify-machineinstrs %s -o - | FileCheck %s
   3 # RUN: llc -march=amdgcn -mcpu=gfx908 -run-pass=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: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
  16     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
  17     ; CHECK: [[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
  18     ; CHECK: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x1f32), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<32 x s32>), 0, 0, 0
  19     ; CHECK: $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>)
  20     %0:_(s32) = COPY $vgpr0
  21     %1:_(s32) = COPY $vgpr1
  22     %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
  23     %3:_(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x1f32), %0, %1, %2, 0, 0, 0
  24     $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
  25 ...
  26
  27 ---
  28 name: mfma_f32_32x32x1f32_sss
  29 legalized: true
  30 tracksRegLiveness: true
  31 body: |
  32   bb.0:
  33     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
  34
  35     ; CHECK-LABEL: name: mfma_f32_32x32x1f32_sss
  36     ; 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
  37     ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
  38     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
  39     ; CHECK: [[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
  40     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
  41     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
  42     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<32 x s32>) = COPY [[COPY2]](<32 x s32>)
  43     ; CHECK: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x1f32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<32 x s32>), 0, 0, 0
  44     ; CHECK: $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>)
  45     %0:_(s32) = COPY $sgpr32
  46     %1:_(s32) = COPY $sgpr33
  47     %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
  48     %3:_(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x1f32), %0, %1, %2, 0, 0, 0
  49     $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
  50 ...
  51
  52 ---
  53 name: mfma_f32_16x16x1f32_vva
  54 legalized: true
  55 tracksRegLiveness: true
  56 body: |
  57   bb.0:
  58     liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
  59
  60     ; CHECK-LABEL: name: mfma_f32_16x16x1f32_vva
  61     ; CHECK: liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
  62     ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
  63     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
  64     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
  65     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x1f32), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<16 x s32>), 0, 0, 0
  66     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
  67     %0:_(s32) = COPY $vgpr0
  68     %1:_(s32) = COPY $vgpr1
  69     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
  70     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x1f32), %0, %1, %2, 0, 0, 0
  71     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
  72 ...
  73
  74 ---
  75 name: mfma_f32_16x16x1f32_sss
  76 legalized: true
  77 tracksRegLiveness: true
  78 body: |
  79   bb.0:
  80     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
  81
  82     ; CHECK-LABEL: name: mfma_f32_16x16x1f32_sss
  83     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
  84     ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
  85     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
  86     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
  87     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
  88     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
  89     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
  90     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x1f32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<16 x s32>), 0, 0, 0
  91     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
  92     %0:_(s32) = COPY $sgpr32
  93     %1:_(s32) = COPY $sgpr33
  94     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
  95     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x1f32), %0, %1, %2, 0, 0, 0
  96     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
  97 ...
  98
  99 ---
 100 name: mfma_f32_4x4x1f32_vva
 101 legalized: true
 102 tracksRegLiveness: true
 103 body: |
 104   bb.0:
 105     liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3
 106
 107     ; CHECK-LABEL: name: mfma_f32_4x4x1f32_vva
 108     ; CHECK: liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3
 109     ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 110     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
 111     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 112     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x1f32), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<4 x s32>), 0, 0, 0
 113     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 114     %0:_(s32) = COPY $vgpr0
 115     %1:_(s32) = COPY $vgpr1
 116     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 117     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x1f32), %0, %1, %2, 0, 0, 0
 118     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 119 ...
 120
 121 ---
 122 name: mfma_f32_4x4x1f32_sss
 123 legalized: true
 124 tracksRegLiveness: true
 125 body: |
 126   bb.0:
 127     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 128
 129     ; CHECK-LABEL: name: mfma_f32_4x4x1f32_sss
 130     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 131     ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 132     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 133     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 134     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 135     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 136     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 137     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x1f32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<4 x s32>), 0, 0, 0
 138     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 139     %0:_(s32) = COPY $sgpr32
 140     %1:_(s32) = COPY $sgpr33
 141     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 142     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x1f32), %0, %1, %2, 0, 0, 0
 143     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 144 ...
 145
 146 ---
 147 name: mfma_f32_32x32x2f32_vva
 148 legalized: true
 149 tracksRegLiveness: true
 150 body: |
 151   bb.0:
 152     liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 153
 154     ; CHECK-LABEL: name: mfma_f32_32x32x2f32_vva
 155     ; CHECK: liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 156     ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 157     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
 158     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 159     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x2f32), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<16 x s32>), 0, 0, 0
 160     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 161     %0:_(s32) = COPY $vgpr0
 162     %1:_(s32) = COPY $vgpr1
 163     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 164     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x2f32), %0, %1, %2, 0, 0, 0
 165     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 166 ...
 167
 168 ---
 169 name: mfma_f32_32x32x2f32_sss
 170 legalized: true
 171 tracksRegLiveness: true
 172 body: |
 173   bb.0:
 174     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 175
 176     ; CHECK-LABEL: name: mfma_f32_32x32x2f32_sss
 177     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 178     ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 179     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 180     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 181     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 182     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 183     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 184     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x2f32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<16 x s32>), 0, 0, 0
 185     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 186     %0:_(s32) = COPY $sgpr32
 187     %1:_(s32) = COPY $sgpr33
 188     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 189     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x2f32), %0, %1, %2, 0, 0, 0
 190     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 191 ...
 192
 193 ---
 194 name: mfma_f32_16x16x4f32_vva
 195 legalized: true
 196 tracksRegLiveness: true
 197 body: |
 198   bb.0:
 199     liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3
 200
 201     ; CHECK-LABEL: name: mfma_f32_16x16x4f32_vva
 202     ; CHECK: liveins: $vgpr0, $vgpr1, $agpr0_agpr1_agpr2_agpr3
 203     ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 204     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
 205     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 206     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f32), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<4 x s32>), 0, 0, 0
 207     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 208     %0:_(s32) = COPY $vgpr0
 209     %1:_(s32) = COPY $vgpr1
 210     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 211     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f32), %0, %1, %2, 0, 0, 0
 212     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 213 ...
 214
 215 ---
 216 name: mfma_f32_16x16x4f32_sss
 217 legalized: true
 218 tracksRegLiveness: true
 219 body: |
 220   bb.0:
 221     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 222
 223     ; CHECK-LABEL: name: mfma_f32_16x16x4f32_sss
 224     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 225     ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 226     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 227     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 228     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 229     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 230     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 231     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<4 x s32>), 0, 0, 0
 232     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 233     %0:_(s32) = COPY $sgpr32
 234     %1:_(s32) = COPY $sgpr33
 235     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 236     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f32), %0, %1, %2, 0, 0, 0
 237     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 238 ...
 239
 240 ---
 241 name: mfma_f32_32x32x4f16_vva
 242 legalized: true
 243 tracksRegLiveness: true
 244 body: |
 245   bb.0:
 246     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
 247
 248     ; CHECK-LABEL: name: mfma_f32_32x32x4f16_vva
 249     ; 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
 250     ; CHECK: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
 251     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
 252     ; CHECK: [[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
 253     ; CHECK: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x4f16), [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<32 x s32>), 0, 0, 0
 254     ; CHECK: $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>)
 255     %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
 256     %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
 257     %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
 258     %3:_(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x4f16), %0, %1, %2, 0, 0, 0
 259     $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
 260 ...
 261
 262 ---
 263 name: mfma_f32_32x32x4f16_sss
 264 legalized: true
 265 tracksRegLiveness: true
 266 body: |
 267   bb.0:
 268     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
 269
 270     ; CHECK-LABEL: name: mfma_f32_32x32x4f16_sss
 271     ; 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
 272     ; CHECK: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr32_sgpr33
 273     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr34_sgpr35
 274     ; CHECK: [[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
 275     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY]](<4 x s16>)
 276     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY1]](<4 x s16>)
 277     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<32 x s32>) = COPY [[COPY2]](<32 x s32>)
 278     ; CHECK: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x4f16), [[COPY3]](<4 x s16>), [[COPY4]](<4 x s16>), [[COPY5]](<32 x s32>), 0, 0, 0
 279     ; CHECK: $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>)
 280     %0:_(<4 x s16>) = COPY $sgpr32_sgpr33
 281     %1:_(<4 x s16>) = COPY $sgpr34_sgpr35
 282     %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
 283     %3:_(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x4f16), %0, %1, %2, 0, 0, 0
 284     $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
 285 ...
 286
 287 ---
 288 name: mfma_f32_16x16x4f16_vva
 289 legalized: true
 290 tracksRegLiveness: true
 291 body: |
 292   bb.0:
 293     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 294
 295     ; CHECK-LABEL: name: mfma_f32_16x16x4f16_vva
 296     ; CHECK: liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 297     ; CHECK: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
 298     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
 299     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 300     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f16), [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<16 x s32>), 0, 0, 0
 301     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 302     %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
 303     %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
 304     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 305     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f16), %0, %1, %2, 0, 0, 0
 306     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 307 ...
 308
 309 ---
 310 name: mfma_f32_16x16x4f16_sss
 311 legalized: true
 312 tracksRegLiveness: true
 313 body: |
 314   bb.0:
 315     liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 316
 317     ; CHECK-LABEL: name: mfma_f32_16x16x4f16_sss
 318     ; CHECK: liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 319     ; CHECK: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr32_sgpr33
 320     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr34_sgpr35
 321     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 322     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY]](<4 x s16>)
 323     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY1]](<4 x s16>)
 324     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 325     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f16), [[COPY3]](<4 x s16>), [[COPY4]](<4 x s16>), [[COPY5]](<16 x s32>), 0, 0, 0
 326     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 327     %0:_(<4 x s16>) = COPY $sgpr32_sgpr33
 328     %1:_(<4 x s16>) = COPY $sgpr34_sgpr35
 329     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 330     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x4f16), %0, %1, %2, 0, 0, 0
 331     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 332 ...
 333
 334 ---
 335 name: mfma_f32_4x4x4f16_vva
 336 legalized: true
 337 tracksRegLiveness: true
 338 body: |
 339   bb.0:
 340     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3
 341
 342     ; CHECK-LABEL: name: mfma_f32_4x4x4f16_vva
 343     ; CHECK: liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3
 344     ; CHECK: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
 345     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
 346     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 347     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x4f16), [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<4 x s32>), 0, 0, 0
 348     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 349     %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
 350     %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
 351     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 352     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x4f16), %0, %1, %2, 0, 0, 0
 353     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 354 ...
 355
 356 ---
 357 name: mfma_f32_4x4x4f16_sss
 358 legalized: true
 359 tracksRegLiveness: true
 360 body: |
 361   bb.0:
 362     liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3
 363
 364     ; CHECK-LABEL: name: mfma_f32_4x4x4f16_sss
 365     ; CHECK: liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3
 366     ; CHECK: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr32_sgpr33
 367     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr34_sgpr35
 368     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 369     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY]](<4 x s16>)
 370     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY1]](<4 x s16>)
 371     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 372     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x4f16), [[COPY3]](<4 x s16>), [[COPY4]](<4 x s16>), [[COPY5]](<4 x s32>), 0, 0, 0
 373     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 374     %0:_(<4 x s16>) = COPY $sgpr32_sgpr33
 375     %1:_(<4 x s16>) = COPY $sgpr34_sgpr35
 376     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 377     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x4f16), %0, %1, %2, 0, 0, 0
 378     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 379 ...
 380
 381 ---
 382 name: mfma_f32_32x32x8f16_vva
 383 legalized: true
 384 tracksRegLiveness: true
 385 body: |
 386   bb.0:
 387     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 388
 389     ; CHECK-LABEL: name: mfma_f32_32x32x8f16_vva
 390     ; CHECK: liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 391     ; CHECK: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
 392     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
 393     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 394     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x8f16), [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<16 x s32>), 0, 0, 0
 395     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 396     %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
 397     %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
 398     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 399     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x8f16), %0, %1, %2, 0, 0, 0
 400     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 401 ...
 402
 403 ---
 404 name: mfma_f32_32x32x8f16_sss
 405 legalized: true
 406 tracksRegLiveness: true
 407 body: |
 408   bb.0:
 409     liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 410
 411     ; CHECK-LABEL: name: mfma_f32_32x32x8f16_sss
 412     ; CHECK: liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 413     ; CHECK: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr32_sgpr33
 414     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr34_sgpr35
 415     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 416     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY]](<4 x s16>)
 417     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY1]](<4 x s16>)
 418     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 419     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x8f16), [[COPY3]](<4 x s16>), [[COPY4]](<4 x s16>), [[COPY5]](<16 x s32>), 0, 0, 0
 420     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 421     %0:_(<4 x s16>) = COPY $sgpr32_sgpr33
 422     %1:_(<4 x s16>) = COPY $sgpr34_sgpr35
 423     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 424     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x8f16), %0, %1, %2, 0, 0, 0
 425     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 426 ...
 427
 428 ---
 429 name: mfma_f32_16x16x16f16_vva
 430 legalized: true
 431 tracksRegLiveness: true
 432 body: |
 433   bb.0:
 434     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3
 435
 436     ; CHECK-LABEL: name: mfma_f32_16x16x16f16_vva
 437     ; CHECK: liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $agpr0_agpr1_agpr2_agpr3
 438     ; CHECK: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
 439     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
 440     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 441     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x16f16), [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<4 x s32>), 0, 0, 0
 442     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 443     %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
 444     %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
 445     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 446     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x16f16), %0, %1, %2, 0, 0, 0
 447     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 448 ...
 449
 450 ---
 451 name: mfma_f32_16x16x16f16_sss
 452 legalized: true
 453 tracksRegLiveness: true
 454 body: |
 455   bb.0:
 456     liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3
 457
 458     ; CHECK-LABEL: name: mfma_f32_16x16x16f16_sss
 459     ; CHECK: liveins: $sgpr32_sgpr33, $sgpr34_sgpr35, $sgpr0_sgpr1_sgpr2_sgpr3
 460     ; CHECK: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr32_sgpr33
 461     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr34_sgpr35
 462     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 463     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY]](<4 x s16>)
 464     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(<4 x s16>) = COPY [[COPY1]](<4 x s16>)
 465     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 466     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x16f16), [[COPY3]](<4 x s16>), [[COPY4]](<4 x s16>), [[COPY5]](<4 x s32>), 0, 0, 0
 467     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 468     %0:_(<4 x s16>) = COPY $sgpr32_sgpr33
 469     %1:_(<4 x s16>) = COPY $sgpr34_sgpr35
 470     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 471     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x16f16), %0, %1, %2, 0, 0, 0
 472     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 473 ...
 474
 475 ---
 476 name: mfma_i32_32x32x4i8_vva
 477 legalized: true
 478 tracksRegLiveness: true
 479 body: |
 480   bb.0:
 481     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
 482
 483     ; CHECK-LABEL: name: mfma_i32_32x32x4i8_vva
 484     ; 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
 485     ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 486     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
 487     ; CHECK: [[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
 488     ; CHECK: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.32x32x4i8), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<32 x s32>), 0, 0, 0
 489     ; CHECK: $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>)
 490     %0:_(s32) = COPY $vgpr0
 491     %1:_(s32) = COPY $vgpr2
 492     %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
 493     %3:_(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.32x32x4i8), %0, %1, %2, 0, 0, 0
 494     $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
 495 ...
 496
 497 ---
 498 name: mfma_i32_32x32x4i8_sss
 499 legalized: true
 500 tracksRegLiveness: true
 501 body: |
 502   bb.0:
 503     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
 504
 505     ; CHECK-LABEL: name: mfma_i32_32x32x4i8_sss
 506     ; 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
 507     ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 508     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 509     ; CHECK: [[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
 510     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 511     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 512     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<32 x s32>) = COPY [[COPY2]](<32 x s32>)
 513     ; CHECK: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.32x32x4i8), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<32 x s32>), 0, 0, 0
 514     ; CHECK: $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>)
 515     %0:_(s32) = COPY $sgpr32
 516     %1:_(s32) = COPY $sgpr33
 517     %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
 518     %3:_(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.32x32x4i8), %0, %1, %2, 0, 0, 0
 519     $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
 520 ...
 521
 522 ---
 523 name: mfma_i32_16x16x4i8_vva
 524 legalized: true
 525 tracksRegLiveness: true
 526 body: |
 527   bb.0:
 528     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 529
 530     ; CHECK-LABEL: name: mfma_i32_16x16x4i8_vva
 531     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 532     ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 533     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
 534     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 535     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.16x16x4i8), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<16 x s32>), 0, 0, 0
 536     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 537     %0:_(s32) = COPY $vgpr0
 538     %1:_(s32) = COPY $vgpr2
 539     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 540     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.16x16x4i8), %0, %1, %2, 0, 0, 0
 541     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 542 ...
 543
 544 ---
 545 name: mfma_i32_16x16x4i8_sss
 546 legalized: true
 547 tracksRegLiveness: true
 548 body: |
 549   bb.0:
 550     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 551
 552     ; CHECK-LABEL: name: mfma_i32_16x16x4i8_sss
 553     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 554     ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 555     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 556     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 557     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 558     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 559     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 560     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.16x16x4i8), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<16 x s32>), 0, 0, 0
 561     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 562     %0:_(s32) = COPY $sgpr32
 563     %1:_(s32) = COPY $sgpr33
 564     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 565     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.16x16x4i8), %0, %1, %2, 0, 0, 0
 566     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 567 ...
 568
 569 ---
 570 name: mfma_i32_4x4x4i8_vva
 571 legalized: true
 572 tracksRegLiveness: true
 573 body: |
 574   bb.0:
 575     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 576
 577     ; CHECK-LABEL: name: mfma_i32_4x4x4i8_vva
 578     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 579     ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 580     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
 581     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 582     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.4x4x4i8), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<4 x s32>), 0, 0, 0
 583     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 584     %0:_(s32) = COPY $vgpr0
 585     %1:_(s32) = COPY $vgpr2
 586     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 587     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.4x4x4i8), %0, %1, %2, 0, 0, 0
 588     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 589 ...
 590
 591 ---
 592 name: mfma_i32_4x4x4i8_sss
 593 legalized: true
 594 tracksRegLiveness: true
 595 body: |
 596   bb.0:
 597     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 598
 599     ; CHECK-LABEL: name: mfma_i32_4x4x4i8_sss
 600     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 601     ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 602     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 603     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 604     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 605     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 606     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 607     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.4x4x4i8), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<4 x s32>), 0, 0, 0
 608     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 609     %0:_(s32) = COPY $sgpr32
 610     %1:_(s32) = COPY $sgpr33
 611     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 612     %3:_(<4 x s32>) = G_INTRINSIC 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_32x32x8i8_vva
 618 legalized: true
 619 tracksRegLiveness: true
 620 body: |
 621   bb.0:
 622     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 623
 624     ; CHECK-LABEL: name: mfma_i32_32x32x8i8_vva
 625     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 626     ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 627     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
 628     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 629     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.32x32x8i8), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<16 x s32>), 0, 0, 0
 630     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 631     %0:_(s32) = COPY $vgpr0
 632     %1:_(s32) = COPY $vgpr2
 633     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 634     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.32x32x8i8), %0, %1, %2, 0, 0, 0
 635     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 636 ...
 637
 638 ---
 639 name: mfma_i32_32x32x8i8_sss
 640 legalized: true
 641 tracksRegLiveness: true
 642 body: |
 643   bb.0:
 644     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 645
 646     ; CHECK-LABEL: name: mfma_i32_32x32x8i8_sss
 647     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 648     ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 649     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 650     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 651     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 652     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 653     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 654     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.32x32x8i8), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<16 x s32>), 0, 0, 0
 655     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 656     %0:_(s32) = COPY $sgpr32
 657     %1:_(s32) = COPY $sgpr33
 658     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 659     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.32x32x8i8), %0, %1, %2, 0, 0, 0
 660     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 661 ...
 662
 663 ---
 664 name: mfma_i32_16x16x16i8_vva
 665 legalized: true
 666 tracksRegLiveness: true
 667 body: |
 668   bb.0:
 669     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 670
 671     ; CHECK-LABEL: name: mfma_i32_16x16x16i8_vva
 672     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 673     ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
 674     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
 675     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 676     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.16x16x16i8), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](<4 x s32>), 0, 0, 0
 677     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 678     %0:_(s32) = COPY $vgpr0
 679     %1:_(s32) = COPY $vgpr2
 680     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 681     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.16x16x16i8), %0, %1, %2, 0, 0, 0
 682     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 683 ...
 684
 685 ---
 686 name: mfma_i32_16x16x16i8_sss
 687 legalized: true
 688 tracksRegLiveness: true
 689 body: |
 690   bb.0:
 691     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 692
 693     ; CHECK-LABEL: name: mfma_i32_16x16x16i8_sss
 694     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 695     ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr32
 696     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr33
 697     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 698     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
 699     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
 700     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 701     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.16x16x16i8), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](<4 x s32>), 0, 0, 0
 702     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 703     %0:_(s32) = COPY $sgpr32
 704     %1:_(s32) = COPY $sgpr33
 705     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 706     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.i32.16x16x16i8), %0, %1, %2, 0, 0, 0
 707     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 708 ...
 709
 710 ---
 711 name: mfma_f32_32x32x2bf16_vva
 712 legalized: true
 713 tracksRegLiveness: true
 714 body: |
 715   bb.0:
 716     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
 717
 718     ; CHECK-LABEL: name: mfma_f32_32x32x2bf16_vva
 719     ; 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
 720     ; CHECK: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
 721     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr2
 722     ; CHECK: [[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
 723     ; CHECK: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x2bf16), [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<32 x s32>), 0, 0, 0
 724     ; CHECK: $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>)
 725     %0:_(<2 x s16>) = COPY $vgpr0
 726     %1:_(<2 x s16>) = COPY $vgpr2
 727     %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
 728     %3:_(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x2bf16), %0, %1, %2, 0, 0, 0
 729     $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
 730 ...
 731
 732 ---
 733 name: mfma_f32_32x32x2bf16_sss
 734 legalized: true
 735 tracksRegLiveness: true
 736 body: |
 737   bb.0:
 738     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
 739
 740     ; CHECK-LABEL: name: mfma_f32_32x32x2bf16_sss
 741     ; 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
 742     ; CHECK: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr32
 743     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr33
 744     ; CHECK: [[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
 745     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY]](<2 x s16>)
 746     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY1]](<2 x s16>)
 747     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<32 x s32>) = COPY [[COPY2]](<32 x s32>)
 748     ; CHECK: [[INT:%[0-9]+]]:agpr(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x2bf16), [[COPY3]](<2 x s16>), [[COPY4]](<2 x s16>), [[COPY5]](<32 x s32>), 0, 0, 0
 749     ; CHECK: $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>)
 750     %0:_(<2 x s16>) = COPY $sgpr32
 751     %1:_(<2 x s16>) = COPY $sgpr33
 752     %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
 753     %3:_(<32 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x2bf16), %0, %1, %2, 0, 0, 0
 754     $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
 755 ...
 756
 757 ---
 758 name: mfma_f32_16x16x2bf16_vva
 759 legalized: true
 760 tracksRegLiveness: true
 761 body: |
 762   bb.0:
 763     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 764
 765     ; CHECK-LABEL: name: mfma_f32_16x16x2bf16_vva
 766     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 767     ; CHECK: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
 768     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr2
 769     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 770     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x2bf16), [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<16 x s32>), 0, 0, 0
 771     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 772     %0:_(<2 x s16>) = COPY $vgpr0
 773     %1:_(<2 x s16>) = COPY $vgpr2
 774     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 775     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x2bf16), %0, %1, %2, 0, 0, 0
 776     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 777 ...
 778
 779 ---
 780 name: mfma_f32_16x16x2bf16_sss
 781 legalized: true
 782 tracksRegLiveness: true
 783 body: |
 784   bb.0:
 785     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 786
 787     ; CHECK-LABEL: name: mfma_f32_16x16x2bf16_sss
 788     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 789     ; CHECK: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr32
 790     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr33
 791     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 792     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY]](<2 x s16>)
 793     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY1]](<2 x s16>)
 794     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 795     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x2bf16), [[COPY3]](<2 x s16>), [[COPY4]](<2 x s16>), [[COPY5]](<16 x s32>), 0, 0, 0
 796     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 797     %0:_(<2 x s16>) = COPY $sgpr32
 798     %1:_(<2 x s16>) = COPY $sgpr33
 799     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 800     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x2bf16), %0, %1, %2, 0, 0, 0
 801     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 802 ...
 803
 804 ---
 805 name: mfma_f32_4x4x2bf16_vva
 806 legalized: true
 807 tracksRegLiveness: true
 808 body: |
 809   bb.0:
 810     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 811
 812     ; CHECK-LABEL: name: mfma_f32_4x4x2bf16_vva
 813     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 814     ; CHECK: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
 815     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr2
 816     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 817     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x2bf16), [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<4 x s32>), 0, 0, 0
 818     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 819     %0:_(<2 x s16>) = COPY $vgpr0
 820     %1:_(<2 x s16>) = COPY $vgpr2
 821     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 822     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x2bf16), %0, %1, %2, 0, 0, 0
 823     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 824 ...
 825
 826 ---
 827 name: mfma_f32_4x4x2bf16_sss
 828 legalized: true
 829 tracksRegLiveness: true
 830 body: |
 831   bb.0:
 832     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 833
 834     ; CHECK-LABEL: name: mfma_f32_4x4x2bf16_sss
 835     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 836     ; CHECK: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr32
 837     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr33
 838     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 839     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY]](<2 x s16>)
 840     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY1]](<2 x s16>)
 841     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 842     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x2bf16), [[COPY3]](<2 x s16>), [[COPY4]](<2 x s16>), [[COPY5]](<4 x s32>), 0, 0, 0
 843     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 844     %0:_(<2 x s16>) = COPY $sgpr32
 845     %1:_(<2 x s16>) = COPY $sgpr33
 846     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 847     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.4x4x2bf16), %0, %1, %2, 0, 0, 0
 848     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 849 ...
 850
 851 ---
 852 name: mfma_f32_32x32x4bf16_vva
 853 legalized: true
 854 tracksRegLiveness: true
 855 body: |
 856   bb.0:
 857     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 858
 859     ; CHECK-LABEL: name: mfma_f32_32x32x4bf16_vva
 860     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 861     ; CHECK: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
 862     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr2
 863     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 864     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x4bf16), [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<16 x s32>), 0, 0, 0
 865     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 866     %0:_(<2 x s16>) = COPY $vgpr0
 867     %1:_(<2 x s16>) = COPY $vgpr2
 868     %2:_(<16 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
 869     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x4bf16), %0, %1, %2, 0, 0, 0
 870     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 871 ...
 872
 873 ---
 874 name: mfma_f32_32x32x4bf16_sss
 875 legalized: true
 876 tracksRegLiveness: true
 877 body: |
 878   bb.0:
 879     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 880
 881     ; CHECK-LABEL: name: mfma_f32_32x32x4bf16_sss
 882     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 883     ; CHECK: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr32
 884     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr33
 885     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 886     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY]](<2 x s16>)
 887     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY1]](<2 x s16>)
 888     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<16 x s32>) = COPY [[COPY2]](<16 x s32>)
 889     ; CHECK: [[INT:%[0-9]+]]:agpr(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x4bf16), [[COPY3]](<2 x s16>), [[COPY4]](<2 x s16>), [[COPY5]](<16 x s32>), 0, 0, 0
 890     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[INT]](<16 x s32>)
 891     %0:_(<2 x s16>) = COPY $sgpr32
 892     %1:_(<2 x s16>) = COPY $sgpr33
 893     %2:_(<16 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
 894     %3:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.32x32x4bf16), %0, %1, %2, 0, 0, 0
 895     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %3
 896 ...
 897
 898 ---
 899 name: mfma_f32_16x16x8bf16_vva
 900 legalized: true
 901 tracksRegLiveness: true
 902 body: |
 903   bb.0:
 904     liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 905
 906     ; CHECK-LABEL: name: mfma_f32_16x16x8bf16_vva
 907     ; CHECK: liveins: $vgpr0, $vgpr2, $agpr0_agpr1_agpr2_agpr3
 908     ; CHECK: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
 909     ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr2
 910     ; CHECK: [[COPY2:%[0-9]+]]:agpr(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 911     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x8bf16), [[COPY]](<2 x s16>), [[COPY1]](<2 x s16>), [[COPY2]](<4 x s32>), 0, 0, 0
 912     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 913     %0:_(<2 x s16>) = COPY $vgpr0
 914     %1:_(<2 x s16>) = COPY $vgpr2
 915     %2:_(<4 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3
 916     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x8bf16), %0, %1, %2, 0, 0, 0
 917     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 918 ...
 919
 920 ---
 921 name: mfma_f32_16x16x8bf16_sss
 922 legalized: true
 923 tracksRegLiveness: true
 924 body: |
 925   bb.0:
 926     liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 927
 928     ; CHECK-LABEL: name: mfma_f32_16x16x8bf16_sss
 929     ; CHECK: liveins: $sgpr32, $sgpr33, $sgpr0_sgpr1_sgpr2_sgpr3
 930     ; CHECK: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr32
 931     ; CHECK: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr33
 932     ; CHECK: [[COPY2:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 933     ; CHECK: [[COPY3:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY]](<2 x s16>)
 934     ; CHECK: [[COPY4:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[COPY1]](<2 x s16>)
 935     ; CHECK: [[COPY5:%[0-9]+]]:agpr(<4 x s32>) = COPY [[COPY2]](<4 x s32>)
 936     ; CHECK: [[INT:%[0-9]+]]:agpr(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x8bf16), [[COPY3]](<2 x s16>), [[COPY4]](<2 x s16>), [[COPY5]](<4 x s32>), 0, 0, 0
 937     ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[INT]](<4 x s32>)
 938     %0:_(<2 x s16>) = COPY $sgpr32
 939     %1:_(<2 x s16>) = COPY $sgpr33
 940     %2:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
 941     %3:_(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.amdgcn.mfma.f32.16x16x8bf16), %0, %1, %2, 0, 0, 0
 942     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3
 943 ...