llvm/test/CodeGen/AMDGPU/fma-mix.gfx11plus.ll

; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -stop-after=amdgpu-isel | FileCheck %s --check-prefixes=GFX11-REAL16
; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -stop-after=amdgpu-isel | FileCheck %s --check-prefixes=GFX11-FAKE16

; Make sure no "vgpr32 = copy vgpr16" is generated

define amdgpu_kernel void @fma_mix_f16 (ptr addrspace(1) %a, ptr addrspace(1) %b, ptr addrspace(1) %c, ptr addrspace(1) %out) {
  ; GFX11-REAL16-LABEL: name: fma_mix_f16
  ; GFX11-REAL16: bb.0.entry:
  ; GFX11-REAL16-NEXT:   liveins: $vgpr0, $sgpr4_sgpr5
  ; GFX11-REAL16-NEXT: {{  $}}
  ; GFX11-REAL16-NEXT:   [[COPY:%[0-9]+]]:sgpr_64(p4) = COPY $sgpr4_sgpr5
  ; GFX11-REAL16-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32(s32) = COPY $vgpr0
  ; GFX11-REAL16-NEXT:   [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[COPY]](p4), 36, 0 :: (dereferenceable invariant load (s256) from %ir.a.kernarg.offset, align 4, addrspace 4)
  ; GFX11-REAL16-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub1
  ; GFX11-REAL16-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub0
  ; GFX11-REAL16-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64_xexec_xnull = REG_SEQUENCE killed [[COPY3]], %subreg.sub0, killed [[COPY2]], %subreg.sub1
  ; GFX11-REAL16-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub3
  ; GFX11-REAL16-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub2
  ; GFX11-REAL16-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec_xnull = REG_SEQUENCE killed [[COPY5]], %subreg.sub0, killed [[COPY4]], %subreg.sub1
  ; GFX11-REAL16-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub5
  ; GFX11-REAL16-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub4
  ; GFX11-REAL16-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec_xnull = REG_SEQUENCE killed [[COPY7]], %subreg.sub0, killed [[COPY6]], %subreg.sub1
  ; GFX11-REAL16-NEXT:   [[COPY8:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub7
  ; GFX11-REAL16-NEXT:   [[COPY9:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub6
  ; GFX11-REAL16-NEXT:   [[REG_SEQUENCE3:%[0-9]+]]:sreg_64_xexec_xnull = REG_SEQUENCE killed [[COPY9]], %subreg.sub0, killed [[COPY8]], %subreg.sub1
  ; GFX11-REAL16-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
  ; GFX11-REAL16-NEXT:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 1023
  ; GFX11-REAL16-NEXT:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY1]](s32), killed [[S_MOV_B32_]], implicit $exec
  ; GFX11-REAL16-NEXT:   [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 2
  ; GFX11-REAL16-NEXT:   [[V_LSHLREV_B32_e64_:%[0-9]+]]:vgpr_32 = nuw nsw V_LSHLREV_B32_e64 killed [[S_MOV_B32_1]], killed [[V_AND_B32_e64_]], implicit $exec
  ; GFX11-REAL16-NEXT:   [[GLOBAL_LOAD_DWORD_SADDR:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_DWORD_SADDR killed [[REG_SEQUENCE]], [[V_LSHLREV_B32_e64_]], 0, 0, implicit $exec :: (load (s32) from %ir.in.gep1, addrspace 1)
  ; GFX11-REAL16-NEXT:   [[GLOBAL_LOAD_DWORD_SADDR1:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_DWORD_SADDR killed [[REG_SEQUENCE1]], [[V_LSHLREV_B32_e64_]], 0, 0, implicit $exec :: (load (s32) from %ir.in.gep2, addrspace 1)
  ; GFX11-REAL16-NEXT:   [[GLOBAL_LOAD_SHORT_D16_SADDR_t16_:%[0-9]+]]:vgpr_16 = GLOBAL_LOAD_SHORT_D16_SADDR_t16 killed [[REG_SEQUENCE2]], [[V_LSHLREV_B32_e64_]], 0, 0, implicit $exec :: (load (s16) from %ir.in.gep3, addrspace 1)
  ; GFX11-REAL16-NEXT:   [[V_MOV_B16_t16_e64_:%[0-9]+]]:vgpr_16 = V_MOV_B16_t16_e64 0, 14336, 0, implicit $exec
  ; GFX11-REAL16-NEXT:   [[V_ADD_F16_t16_e64_:%[0-9]+]]:vgpr_16 = nofpexcept V_ADD_F16_t16_e64 0, killed [[GLOBAL_LOAD_SHORT_D16_SADDR_t16_]], 0, killed [[V_MOV_B16_t16_e64_]], 0, 0, 0, implicit $mode, implicit $exec
  ; GFX11-REAL16-NEXT:   [[DEF:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
  ; GFX11-REAL16-NEXT:   [[DEF1:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
  ; GFX11-REAL16-NEXT:   [[REG_SEQUENCE4:%[0-9]+]]:vgpr_32 = REG_SEQUENCE killed [[V_ADD_F16_t16_e64_]], %subreg.lo16, killed [[DEF]], %subreg.hi16
  ; GFX11-REAL16-NEXT:   [[V_FMA_MIX_F16_t16_:%[0-9]+]]:vgpr_16 = nofpexcept V_FMA_MIX_F16_t16 0, killed [[GLOBAL_LOAD_DWORD_SADDR]], 0, killed [[GLOBAL_LOAD_DWORD_SADDR1]], 8, killed [[REG_SEQUENCE4]], 0, 0, 0, implicit $mode, implicit $exec
  ; GFX11-REAL16-NEXT:   GLOBAL_STORE_SHORT_SADDR_t16 killed [[V_MOV_B32_e32_]], killed [[V_FMA_MIX_F16_t16_]], killed [[REG_SEQUENCE3]], 0, 0, implicit $exec :: (store (s16) into %ir.4, addrspace 1)
  ; GFX11-REAL16-NEXT:   S_ENDPGM 0
  ;
  ; GFX11-FAKE16-LABEL: name: fma_mix_f16
  ; GFX11-FAKE16: bb.0.entry:
  ; GFX11-FAKE16-NEXT:   liveins: $vgpr0, $sgpr4_sgpr5
  ; GFX11-FAKE16-NEXT: {{  $}}
  ; GFX11-FAKE16-NEXT:   [[COPY:%[0-9]+]]:sgpr_64(p4) = COPY $sgpr4_sgpr5
  ; GFX11-FAKE16-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32(s32) = COPY $vgpr0
  ; GFX11-FAKE16-NEXT:   [[S_LOAD_DWORDX8_IMM:%[0-9]+]]:sgpr_256 = S_LOAD_DWORDX8_IMM [[COPY]](p4), 36, 0 :: (dereferenceable invariant load (s256) from %ir.a.kernarg.offset, align 4, addrspace 4)
  ; GFX11-FAKE16-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub1
  ; GFX11-FAKE16-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub0
  ; GFX11-FAKE16-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64_xexec_xnull = REG_SEQUENCE killed [[COPY3]], %subreg.sub0, killed [[COPY2]], %subreg.sub1
  ; GFX11-FAKE16-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub3
  ; GFX11-FAKE16-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub2
  ; GFX11-FAKE16-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec_xnull = REG_SEQUENCE killed [[COPY5]], %subreg.sub0, killed [[COPY4]], %subreg.sub1
  ; GFX11-FAKE16-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub5
  ; GFX11-FAKE16-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub4
  ; GFX11-FAKE16-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec_xnull = REG_SEQUENCE killed [[COPY7]], %subreg.sub0, killed [[COPY6]], %subreg.sub1
  ; GFX11-FAKE16-NEXT:   [[COPY8:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub7
  ; GFX11-FAKE16-NEXT:   [[COPY9:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX8_IMM]].sub6
  ; GFX11-FAKE16-NEXT:   [[REG_SEQUENCE3:%[0-9]+]]:sreg_64_xexec_xnull = REG_SEQUENCE killed [[COPY9]], %subreg.sub0, killed [[COPY8]], %subreg.sub1
  ; GFX11-FAKE16-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
  ; GFX11-FAKE16-NEXT:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 1023
  ; GFX11-FAKE16-NEXT:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY1]](s32), killed [[S_MOV_B32_]], implicit $exec
  ; GFX11-FAKE16-NEXT:   [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 2
  ; GFX11-FAKE16-NEXT:   [[V_LSHLREV_B32_e64_:%[0-9]+]]:vgpr_32 = nuw nsw V_LSHLREV_B32_e64 killed [[S_MOV_B32_1]], killed [[V_AND_B32_e64_]], implicit $exec
  ; GFX11-FAKE16-NEXT:   [[GLOBAL_LOAD_DWORD_SADDR:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_DWORD_SADDR killed [[REG_SEQUENCE]], [[V_LSHLREV_B32_e64_]], 0, 0, implicit $exec :: (load (s32) from %ir.in.gep1, addrspace 1)
  ; GFX11-FAKE16-NEXT:   [[GLOBAL_LOAD_DWORD_SADDR1:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_DWORD_SADDR killed [[REG_SEQUENCE1]], [[V_LSHLREV_B32_e64_]], 0, 0, implicit $exec :: (load (s32) from %ir.in.gep2, addrspace 1)
  ; GFX11-FAKE16-NEXT:   [[GLOBAL_LOAD_USHORT_SADDR:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_USHORT_SADDR killed [[REG_SEQUENCE2]], [[V_LSHLREV_B32_e64_]], 0, 0, implicit $exec :: (load (s16) from %ir.in.gep3, addrspace 1)
  ; GFX11-FAKE16-NEXT:   [[S_MOV_B32_2:%[0-9]+]]:sreg_32 = S_MOV_B32 14336
  ; GFX11-FAKE16-NEXT:   [[V_ADD_F16_fake16_e64_:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F16_fake16_e64 0, killed [[GLOBAL_LOAD_USHORT_SADDR]], 0, killed [[S_MOV_B32_2]], 0, 0, implicit $mode, implicit $exec
  ; GFX11-FAKE16-NEXT:   [[DEF:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
  ; GFX11-FAKE16-NEXT:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[DEF]]
  ; GFX11-FAKE16-NEXT:   [[V_FMA_MIXLO_F16_:%[0-9]+]]:vgpr_32 = nofpexcept V_FMA_MIXLO_F16 0, killed [[GLOBAL_LOAD_DWORD_SADDR]], 0, killed [[GLOBAL_LOAD_DWORD_SADDR1]], 8, killed [[V_ADD_F16_fake16_e64_]], 0, [[COPY10]], 0, 0, implicit $mode, implicit $exec
  ; GFX11-FAKE16-NEXT:   GLOBAL_STORE_SHORT_SADDR killed [[V_MOV_B32_e32_]], killed [[V_FMA_MIXLO_F16_]], killed [[REG_SEQUENCE3]], 0, 0, implicit $exec :: (store (s16) into %ir.4, addrspace 1)
  ; GFX11-FAKE16-NEXT:   S_ENDPGM 0
entry:
  %tid = call i32 @llvm.amdgcn.workitem.id.x()
  %in.gep1 = getelementptr i32, ptr addrspace(1) %a, i32 %tid
  %in.gep2 = getelementptr i32, ptr addrspace(1) %b, i32 %tid
  %in.gep3 = getelementptr i32, ptr addrspace(1) %c, i32 %tid
  %load.a = load float, ptr addrspace(1) %in.gep1
  %load.b = load float, ptr addrspace(1) %in.gep2
  %load.c = load half, ptr addrspace(1) %in.gep3
  %add.c = fadd half %load.c, 0.5
  %load.float.c = fpext half %add.c to float
  %result = tail call float @llvm.fmuladd.f32(float %load.a, float %load.b, float %load.float.c)
  %half = fptrunc float %result to half
  store half %half, ptr addrspace(1) %out
  ret void
}