src/llvm-project/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.cvt.fp8.dpp.mir

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs -run-pass=gcn-dpp-combine %s -o - | FileCheck -check-prefix=GFX12 %s

---
name:            test_cvt_f32_bf8_byte0
tracksRegLiveness: true
body:             |
  bb.0:
    liveins: $vgpr0

    ; GFX12-LABEL: name: test_cvt_f32_bf8_byte0
    ; GFX12: liveins: $vgpr0
    ; GFX12-NEXT: {{  $}}
    ; GFX12-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
    ; GFX12-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
    ; GFX12-NEXT: [[V_CVT_F32_BF8_dpp:%[0-9]+]]:vgpr_32 = V_CVT_F32_BF8_dpp [[DEF]], [[COPY]], 228, 15, 15, 1, implicit $mode, implicit $exec
    ; GFX12-NEXT: $vgpr0 = COPY [[V_CVT_F32_BF8_dpp]]
    ; GFX12-NEXT: SI_RETURN_TO_EPILOG $vgpr0
    %0:vgpr_32 = COPY $vgpr0
    %1:vgpr_32 = V_MOV_B32_dpp %0, %0, 228, 15, 15, -1, implicit $exec
    %2:vgpr_32 = V_CVT_F32_BF8_e32 killed %1, implicit $mode, implicit $exec
    $vgpr0 = COPY %2
    SI_RETURN_TO_EPILOG $vgpr0

...
---
name:            test_cvt_f32_bf8_byte2
tracksRegLiveness: true
body:             |
  bb.0:
    liveins: $vgpr0

    ; GFX12-LABEL: name: test_cvt_f32_bf8_byte2
    ; GFX12: liveins: $vgpr0
    ; GFX12-NEXT: {{  $}}
    ; GFX12-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
    ; GFX12-NEXT: [[V_MOV_B32_dpp:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY]], [[COPY]], 228, 15, 15, -1, implicit $exec
    ; GFX12-NEXT: [[V_CVT_F32_BF8_OP_SEL_e64_:%[0-9]+]]:vgpr_32 = V_CVT_F32_BF8_OP_SEL_e64 8, killed [[V_MOV_B32_dpp]], 0, implicit $mode, implicit $exec
    ; GFX12-NEXT: $vgpr0 = COPY [[V_CVT_F32_BF8_OP_SEL_e64_]]
    ; GFX12-NEXT: SI_RETURN_TO_EPILOG $vgpr0
    %0:vgpr_32 = COPY $vgpr0
    %1:vgpr_32 = V_MOV_B32_dpp %0, %0, 228, 15, 15, -1, implicit $exec
    %2:vgpr_32 = V_CVT_F32_BF8_OP_SEL_e64 8, killed %1, 0, implicit $mode, implicit $exec
    $vgpr0 = COPY %2
    SI_RETURN_TO_EPILOG $vgpr0

...
---
name:            test_cvt_f32_fp8_byte3
tracksRegLiveness: true
body:             |
  bb.0:
    liveins: $vgpr0

    ; GFX12-LABEL: name: test_cvt_f32_fp8_byte3
    ; GFX12: liveins: $vgpr0
    ; GFX12-NEXT: {{  $}}
    ; GFX12-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
    ; GFX12-NEXT: [[V_MOV_B32_dpp:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY]], [[COPY]], 228, 15, 15, -1, implicit $exec
    ; GFX12-NEXT: [[V_CVT_F32_FP8_OP_SEL_e64_:%[0-9]+]]:vgpr_32 = V_CVT_F32_FP8_OP_SEL_e64 12, killed [[V_MOV_B32_dpp]], 0, implicit $mode, implicit $exec
    ; GFX12-NEXT: $vgpr0 = COPY [[V_CVT_F32_FP8_OP_SEL_e64_]]
    ; GFX12-NEXT: SI_RETURN_TO_EPILOG $vgpr0
    %0:vgpr_32 = COPY $vgpr0
    %1:vgpr_32 = V_MOV_B32_dpp %0, %0, 228, 15, 15, -1, implicit $exec
    %2:vgpr_32 = V_CVT_F32_FP8_OP_SEL_e64 12, killed %1, 0, implicit $mode, implicit $exec
    $vgpr0 = COPY %2
    SI_RETURN_TO_EPILOG $vgpr0

...
---
name:            test_cvt_pk_bf8_f32_word0
tracksRegLiveness: true
body:             |
  bb.0:
    liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4

    ; GFX12-LABEL: name: test_cvt_pk_bf8_f32_word0
    ; GFX12: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4
    ; GFX12-NEXT: {{  $}}
    ; GFX12-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr4
    ; GFX12-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr3
    ; GFX12-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
    ; GFX12-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr1
    ; GFX12-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
    ; GFX12-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
    ; GFX12-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
    ; GFX12-NEXT: [[V_CVT_PK_BF8_F32_e64_dpp:%[0-9]+]]:vgpr_32 = V_CVT_PK_BF8_F32_e64_dpp [[DEF]], 0, [[COPY4]], 0, [[COPY3]], [[COPY2]], 0, 228, 15, 15, 1, implicit $mode, implicit $exec
    ; GFX12-NEXT: GLOBAL_STORE_DWORD [[REG_SEQUENCE]], killed [[V_CVT_PK_BF8_F32_e64_dpp]], 0, 0, implicit $exec
    ; GFX12-NEXT: S_ENDPGM 0
    %4:vgpr_32 = COPY $vgpr4
    %3:vgpr_32 = COPY $vgpr3
    %2:vgpr_32 = COPY $vgpr2
    %1:vgpr_32 = COPY $vgpr1
    %0:vgpr_32 = COPY $vgpr0
    %11:vreg_64 = REG_SEQUENCE %3, %subreg.sub0, %4, %subreg.sub1
    %6:vgpr_32 = V_MOV_B32_dpp %0, %0, 228, 15, 15, -1, implicit $exec
    %7:vgpr_32 = V_CVT_PK_BF8_F32_e64 0, killed %6, 0, %1, %2, 0, implicit $mode, implicit $exec
    GLOBAL_STORE_DWORD %11, killed %7, 0, 0, implicit $exec
    S_ENDPGM 0

...
---
name:            test_cvt_pk_fp8_f32_word1
tracksRegLiveness: true
body:             |
  bb.0:
    liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4

    ; GFX12-LABEL: name: test_cvt_pk_fp8_f32_word1
    ; GFX12: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4
    ; GFX12-NEXT: {{  $}}
    ; GFX12-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr4
    ; GFX12-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr3
    ; GFX12-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
    ; GFX12-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr1
    ; GFX12-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
    ; GFX12-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
    ; GFX12-NEXT: [[V_MOV_B32_dpp:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY4]], [[COPY4]], 228, 15, 15, -1, implicit $exec
    ; GFX12-NEXT: [[V_CVT_PK_FP8_F32_e64_:%[0-9]+]]:vgpr_32 = V_CVT_PK_FP8_F32_e64 8, killed [[V_MOV_B32_dpp]], 0, [[COPY3]], [[COPY2]], 0, implicit $mode, implicit $exec
    ; GFX12-NEXT: GLOBAL_STORE_DWORD [[REG_SEQUENCE]], killed [[V_CVT_PK_FP8_F32_e64_]], 0, 0, implicit $exec
    ; GFX12-NEXT: S_ENDPGM 0
    %4:vgpr_32 = COPY $vgpr4
    %3:vgpr_32 = COPY $vgpr3
    %2:vgpr_32 = COPY $vgpr2
    %1:vgpr_32 = COPY $vgpr1
    %0:vgpr_32 = COPY $vgpr0
    %11:vreg_64 = REG_SEQUENCE %3, %subreg.sub0, %4, %subreg.sub1
    %6:vgpr_32 = V_MOV_B32_dpp %0, %0, 228, 15, 15, -1, implicit $exec
    %7:vgpr_32 = V_CVT_PK_FP8_F32_e64 8, killed %6, 0, %1, %2, 0, implicit $mode, implicit $exec
    GLOBAL_STORE_DWORD %11, killed %7, 0, 0, implicit $exec
    S_ENDPGM 0

...
---
name:            test_cvt_sr_bf8_f32_byte0
tracksRegLiveness: true
body:             |
  bb.0:
    liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4

    ; GFX12-LABEL: name: test_cvt_sr_bf8_f32_byte0
    ; GFX12: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4
    ; GFX12-NEXT: {{  $}}
    ; GFX12-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr4
    ; GFX12-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr3
    ; GFX12-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
    ; GFX12-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr1
    ; GFX12-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
    ; GFX12-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
    ; GFX12-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
    ; GFX12-NEXT: [[V_CVT_SR_BF8_F32_e64_dpp:%[0-9]+]]:vgpr_32 = V_CVT_SR_BF8_F32_e64_dpp [[DEF]], 0, [[COPY4]], 0, [[COPY3]], 0, [[COPY2]], 0, 228, 15, 15, 1, implicit $mode, implicit $exec
    ; GFX12-NEXT: GLOBAL_STORE_DWORD [[REG_SEQUENCE]], killed [[V_CVT_SR_BF8_F32_e64_dpp]], 0, 0, implicit $exec
    ; GFX12-NEXT: S_ENDPGM 0
    %4:vgpr_32 = COPY $vgpr4
    %3:vgpr_32 = COPY $vgpr3
    %2:vgpr_32 = COPY $vgpr2
    %1:vgpr_32 = COPY $vgpr1
    %0:vgpr_32 = COPY $vgpr0
    %11:vreg_64 = REG_SEQUENCE %3, %subreg.sub0, %4, %subreg.sub1
    %6:vgpr_32 = V_MOV_B32_dpp %0, %0, 228, 15, 15, -1, implicit $exec
    %7:vgpr_32 = V_CVT_SR_BF8_F32_e64 0, killed %6, 0, %1, 0, %2, 0, implicit $mode, implicit $exec
    GLOBAL_STORE_DWORD %11, killed %7, 0, 0, implicit $exec
    S_ENDPGM 0

...
---
name:            test_cvt_sr_fp8_f32_byte2
tracksRegLiveness: true
body:             |
  bb.0:
    liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4

    ; GFX12-LABEL: name: test_cvt_sr_fp8_f32_byte2
    ; GFX12: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4
    ; GFX12-NEXT: {{  $}}
    ; GFX12-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr4
    ; GFX12-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr3
    ; GFX12-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
    ; GFX12-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr1
    ; GFX12-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
    ; GFX12-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
    ; GFX12-NEXT: [[V_MOV_B32_dpp:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY4]], [[COPY4]], 228, 15, 15, -1, implicit $exec
    ; GFX12-NEXT: [[V_CVT_SR_FP8_F32_e64_:%[0-9]+]]:vgpr_32 = V_CVT_SR_FP8_F32_e64 8, killed [[V_MOV_B32_dpp]], 0, [[COPY3]], 0, [[COPY2]], 0, implicit $mode, implicit $exec
    ; GFX12-NEXT: GLOBAL_STORE_DWORD [[REG_SEQUENCE]], killed [[V_CVT_SR_FP8_F32_e64_]], 0, 0, implicit $exec
    ; GFX12-NEXT: S_ENDPGM 0
    %4:vgpr_32 = COPY $vgpr4
    %3:vgpr_32 = COPY $vgpr3
    %2:vgpr_32 = COPY $vgpr2
    %1:vgpr_32 = COPY $vgpr1
    %0:vgpr_32 = COPY $vgpr0
    %11:vreg_64 = REG_SEQUENCE %3, %subreg.sub0, %4, %subreg.sub1
    %6:vgpr_32 = V_MOV_B32_dpp %0, %0, 228, 15, 15, -1, implicit $exec
    %7:vgpr_32 = V_CVT_SR_FP8_F32_e64 8, killed %6, 0, %1, 0, %2, 0, implicit $mode, implicit $exec
    GLOBAL_STORE_DWORD %11, killed %7, 0, 0, implicit $exec
    S_ENDPGM 0

...