test/CodeGen/X86/fp-strict-scalar-fp16.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2  -O3 | FileCheck %s --check-prefixes=SSE2
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+f16c  -O3 | FileCheck %s --check-prefixes=AVX,F16C
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f  -O3 | FileCheck %s --check-prefixes=AVX,AVX512
; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512fp16 -mattr=+avx512vl -O3 | FileCheck %s --check-prefixes=X86
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512fp16 -mattr=+avx512vl -O3 | FileCheck %s --check-prefixes=X64

declare half @llvm.experimental.constrained.fadd.f16(half, half, metadata, metadata)
declare half @llvm.experimental.constrained.fsub.f16(half, half, metadata, metadata)
declare half @llvm.experimental.constrained.fmul.f16(half, half, metadata, metadata)
declare half @llvm.experimental.constrained.fdiv.f16(half, half, metadata, metadata)
declare float @llvm.experimental.constrained.fpext.f32.f16(half, metadata)
declare double @llvm.experimental.constrained.fpext.f64.f16(half, metadata)
declare half @llvm.experimental.constrained.fptrunc.f16.f32(float, metadata, metadata)
declare half @llvm.experimental.constrained.fptrunc.f16.f64(double, metadata, metadata)
declare half @llvm.experimental.constrained.sqrt.f16(half, metadata, metadata)
declare half @llvm.experimental.constrained.fma.f16(half, half, half, metadata, metadata)

define half @fadd_f16(half %a, half %b) nounwind strictfp {
; SSE2-LABEL: fadd_f16:
; SSE2:       # %bb.0:
; SSE2-NEXT:    pushq %rax
; SSE2-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; SSE2-NEXT:    movaps %xmm1, %xmm0
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    movss %xmm0, (%rsp) # 4-byte Spill
; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
; SSE2-NEXT:    # xmm0 = mem[0],zero,zero,zero
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    addss (%rsp), %xmm0 # 4-byte Folded Reload
; SSE2-NEXT:    callq __truncsfhf2@PLT
; SSE2-NEXT:    popq %rax
; SSE2-NEXT:    retq
;
; AVX-LABEL: fadd_f16:
; AVX:       # %bb.0:
; AVX-NEXT:    vpextrw $0, %xmm0, %eax
; AVX-NEXT:    vpextrw $0, %xmm1, %ecx
; AVX-NEXT:    movzwl %cx, %ecx
; AVX-NEXT:    vmovd %ecx, %xmm0
; AVX-NEXT:    vcvtph2ps %xmm0, %xmm0
; AVX-NEXT:    movzwl %ax, %eax
; AVX-NEXT:    vmovd %eax, %xmm1
; AVX-NEXT:    vcvtph2ps %xmm1, %xmm1
; AVX-NEXT:    vaddss %xmm0, %xmm1, %xmm0
; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
; AVX-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
; AVX-NEXT:    vmovd %xmm0, %eax
; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
; AVX-NEXT:    retq
;
; X86-LABEL: fadd_f16:
; X86:       # %bb.0:
; X86-NEXT:    vmovsh {{[0-9]+}}(%esp), %xmm0
; X86-NEXT:    vaddsh {{[0-9]+}}(%esp), %xmm0, %xmm0
; X86-NEXT:    retl
;
; X64-LABEL: fadd_f16:
; X64:       # %bb.0:
; X64-NEXT:    vaddsh %xmm1, %xmm0, %xmm0
; X64-NEXT:    retq
  %ret = call half @llvm.experimental.constrained.fadd.f16(half %a, half %b,
                                                           metadata !"round.dynamic",
                                                           metadata !"fpexcept.strict") #0
  ret half %ret
}

define half @fsub_f16(half %a, half %b) nounwind strictfp {
; SSE2-LABEL: fsub_f16:
; SSE2:       # %bb.0:
; SSE2-NEXT:    pushq %rax
; SSE2-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; SSE2-NEXT:    movaps %xmm1, %xmm0
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    movss %xmm0, (%rsp) # 4-byte Spill
; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
; SSE2-NEXT:    # xmm0 = mem[0],zero,zero,zero
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    subss (%rsp), %xmm0 # 4-byte Folded Reload
; SSE2-NEXT:    callq __truncsfhf2@PLT
; SSE2-NEXT:    popq %rax
; SSE2-NEXT:    retq
;
; AVX-LABEL: fsub_f16:
; AVX:       # %bb.0:
; AVX-NEXT:    vpextrw $0, %xmm0, %eax
; AVX-NEXT:    vpextrw $0, %xmm1, %ecx
; AVX-NEXT:    movzwl %cx, %ecx
; AVX-NEXT:    vmovd %ecx, %xmm0
; AVX-NEXT:    vcvtph2ps %xmm0, %xmm0
; AVX-NEXT:    movzwl %ax, %eax
; AVX-NEXT:    vmovd %eax, %xmm1
; AVX-NEXT:    vcvtph2ps %xmm1, %xmm1
; AVX-NEXT:    vsubss %xmm0, %xmm1, %xmm0
; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
; AVX-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
; AVX-NEXT:    vmovd %xmm0, %eax
; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
; AVX-NEXT:    retq
;
; X86-LABEL: fsub_f16:
; X86:       # %bb.0:
; X86-NEXT:    vmovsh {{[0-9]+}}(%esp), %xmm0
; X86-NEXT:    vsubsh {{[0-9]+}}(%esp), %xmm0, %xmm0
; X86-NEXT:    retl
;
; X64-LABEL: fsub_f16:
; X64:       # %bb.0:
; X64-NEXT:    vsubsh %xmm1, %xmm0, %xmm0
; X64-NEXT:    retq
  %ret = call half @llvm.experimental.constrained.fsub.f16(half %a, half %b,
                                                           metadata !"round.dynamic",
                                                           metadata !"fpexcept.strict") #0
  ret half %ret
}

define half @fmul_f16(half %a, half %b) nounwind strictfp {
; SSE2-LABEL: fmul_f16:
; SSE2:       # %bb.0:
; SSE2-NEXT:    pushq %rax
; SSE2-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; SSE2-NEXT:    movaps %xmm1, %xmm0
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    movss %xmm0, (%rsp) # 4-byte Spill
; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
; SSE2-NEXT:    # xmm0 = mem[0],zero,zero,zero
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    mulss (%rsp), %xmm0 # 4-byte Folded Reload
; SSE2-NEXT:    callq __truncsfhf2@PLT
; SSE2-NEXT:    popq %rax
; SSE2-NEXT:    retq
;
; AVX-LABEL: fmul_f16:
; AVX:       # %bb.0:
; AVX-NEXT:    vpextrw $0, %xmm0, %eax
; AVX-NEXT:    vpextrw $0, %xmm1, %ecx
; AVX-NEXT:    movzwl %cx, %ecx
; AVX-NEXT:    vmovd %ecx, %xmm0
; AVX-NEXT:    vcvtph2ps %xmm0, %xmm0
; AVX-NEXT:    movzwl %ax, %eax
; AVX-NEXT:    vmovd %eax, %xmm1
; AVX-NEXT:    vcvtph2ps %xmm1, %xmm1
; AVX-NEXT:    vmulss %xmm0, %xmm1, %xmm0
; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
; AVX-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
; AVX-NEXT:    vmovd %xmm0, %eax
; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
; AVX-NEXT:    retq
;
; X86-LABEL: fmul_f16:
; X86:       # %bb.0:
; X86-NEXT:    vmovsh {{[0-9]+}}(%esp), %xmm0
; X86-NEXT:    vmulsh {{[0-9]+}}(%esp), %xmm0, %xmm0
; X86-NEXT:    retl
;
; X64-LABEL: fmul_f16:
; X64:       # %bb.0:
; X64-NEXT:    vmulsh %xmm1, %xmm0, %xmm0
; X64-NEXT:    retq
  %ret = call half @llvm.experimental.constrained.fmul.f16(half %a, half %b,
                                                           metadata !"round.dynamic",
                                                           metadata !"fpexcept.strict") #0
  ret half %ret
}

define half @fdiv_f16(half %a, half %b) nounwind strictfp {
; SSE2-LABEL: fdiv_f16:
; SSE2:       # %bb.0:
; SSE2-NEXT:    pushq %rax
; SSE2-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; SSE2-NEXT:    movaps %xmm1, %xmm0
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    movss %xmm0, (%rsp) # 4-byte Spill
; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
; SSE2-NEXT:    # xmm0 = mem[0],zero,zero,zero
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    divss (%rsp), %xmm0 # 4-byte Folded Reload
; SSE2-NEXT:    callq __truncsfhf2@PLT
; SSE2-NEXT:    popq %rax
; SSE2-NEXT:    retq
;
; AVX-LABEL: fdiv_f16:
; AVX:       # %bb.0:
; AVX-NEXT:    vpextrw $0, %xmm0, %eax
; AVX-NEXT:    vpextrw $0, %xmm1, %ecx
; AVX-NEXT:    movzwl %cx, %ecx
; AVX-NEXT:    vmovd %ecx, %xmm0
; AVX-NEXT:    vcvtph2ps %xmm0, %xmm0
; AVX-NEXT:    movzwl %ax, %eax
; AVX-NEXT:    vmovd %eax, %xmm1
; AVX-NEXT:    vcvtph2ps %xmm1, %xmm1
; AVX-NEXT:    vdivss %xmm0, %xmm1, %xmm0
; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
; AVX-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
; AVX-NEXT:    vmovd %xmm0, %eax
; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
; AVX-NEXT:    retq
;
; X86-LABEL: fdiv_f16:
; X86:       # %bb.0:
; X86-NEXT:    vmovsh {{[0-9]+}}(%esp), %xmm0
; X86-NEXT:    vdivsh {{[0-9]+}}(%esp), %xmm0, %xmm0
; X86-NEXT:    retl
;
; X64-LABEL: fdiv_f16:
; X64:       # %bb.0:
; X64-NEXT:    vdivsh %xmm1, %xmm0, %xmm0
; X64-NEXT:    retq
  %ret = call half @llvm.experimental.constrained.fdiv.f16(half %a, half %b,
                                                           metadata !"round.dynamic",
                                                           metadata !"fpexcept.strict") #0
  ret half %ret
}

define void @fpext_f16_to_f32(ptr %val, ptr %ret) nounwind strictfp {
; SSE2-LABEL: fpext_f16_to_f32:
; SSE2:       # %bb.0:
; SSE2-NEXT:    pushq %rbx
; SSE2-NEXT:    movq %rsi, %rbx
; SSE2-NEXT:    pinsrw $0, (%rdi), %xmm0
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    movd %xmm0, (%rbx)
; SSE2-NEXT:    popq %rbx
; SSE2-NEXT:    retq
;
; AVX-LABEL: fpext_f16_to_f32:
; AVX:       # %bb.0:
; AVX-NEXT:    movzwl (%rdi), %eax
; AVX-NEXT:    vmovd %eax, %xmm0
; AVX-NEXT:    vcvtph2ps %xmm0, %xmm0
; AVX-NEXT:    vmovss %xmm0, (%rsi)
; AVX-NEXT:    retq
;
; X86-LABEL: fpext_f16_to_f32:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    vmovsh (%ecx), %xmm0
; X86-NEXT:    vcvtsh2ss %xmm0, %xmm0, %xmm0
; X86-NEXT:    vmovss %xmm0, (%eax)
; X86-NEXT:    retl
;
; X64-LABEL: fpext_f16_to_f32:
; X64:       # %bb.0:
; X64-NEXT:    vmovsh (%rdi), %xmm0
; X64-NEXT:    vcvtsh2ss %xmm0, %xmm0, %xmm0
; X64-NEXT:    vmovss %xmm0, (%rsi)
; X64-NEXT:    retq
  %1 = load half, ptr %val, align 4
  %res = call float @llvm.experimental.constrained.fpext.f32.f16(half %1,
                                                                 metadata !"fpexcept.strict") #0
  store float %res, ptr %ret, align 8
  ret void
}

define void @fpext_f16_to_f64(ptr %val, ptr %ret) nounwind strictfp {
; SSE2-LABEL: fpext_f16_to_f64:
; SSE2:       # %bb.0:
; SSE2-NEXT:    pushq %rbx
; SSE2-NEXT:    movq %rsi, %rbx
; SSE2-NEXT:    pinsrw $0, (%rdi), %xmm0
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    cvtss2sd %xmm0, %xmm0
; SSE2-NEXT:    movsd %xmm0, (%rbx)
; SSE2-NEXT:    popq %rbx
; SSE2-NEXT:    retq
;
; AVX-LABEL: fpext_f16_to_f64:
; AVX:       # %bb.0:
; AVX-NEXT:    movzwl (%rdi), %eax
; AVX-NEXT:    vmovd %eax, %xmm0
; AVX-NEXT:    vcvtph2ps %xmm0, %xmm0
; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
; AVX-NEXT:    vmovsd %xmm0, (%rsi)
; AVX-NEXT:    retq
;
; X86-LABEL: fpext_f16_to_f64:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    vmovsh (%ecx), %xmm0
; X86-NEXT:    vcvtsh2sd %xmm0, %xmm0, %xmm0
; X86-NEXT:    vmovsd %xmm0, (%eax)
; X86-NEXT:    retl
;
; X64-LABEL: fpext_f16_to_f64:
; X64:       # %bb.0:
; X64-NEXT:    vmovsh (%rdi), %xmm0
; X64-NEXT:    vcvtsh2sd %xmm0, %xmm0, %xmm0
; X64-NEXT:    vmovsd %xmm0, (%rsi)
; X64-NEXT:    retq
  %1 = load half, ptr %val, align 4
  %res = call double @llvm.experimental.constrained.fpext.f64.f16(half %1,
                                                                  metadata !"fpexcept.strict") #0
  store double %res, ptr %ret, align 8
  ret void
}

define void @fptrunc_float_to_f16(ptr %val, ptr%ret) nounwind strictfp {
; SSE2-LABEL: fptrunc_float_to_f16:
; SSE2:       # %bb.0:
; SSE2-NEXT:    pushq %rbx
; SSE2-NEXT:    movq %rsi, %rbx
; SSE2-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
; SSE2-NEXT:    callq __truncsfhf2@PLT
; SSE2-NEXT:    pextrw $0, %xmm0, %eax
; SSE2-NEXT:    movw %ax, (%rbx)
; SSE2-NEXT:    popq %rbx
; SSE2-NEXT:    retq
;
; AVX-LABEL: fptrunc_float_to_f16:
; AVX:       # %bb.0:
; AVX-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; AVX-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
; AVX-NEXT:    vmovd %xmm0, %eax
; AVX-NEXT:    movw %ax, (%rsi)
; AVX-NEXT:    retq
;
; X86-LABEL: fptrunc_float_to_f16:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; X86-NEXT:    vcvtss2sh %xmm0, %xmm0, %xmm0
; X86-NEXT:    vmovsh %xmm0, (%eax)
; X86-NEXT:    retl
;
; X64-LABEL: fptrunc_float_to_f16:
; X64:       # %bb.0:
; X64-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; X64-NEXT:    vcvtss2sh %xmm0, %xmm0, %xmm0
; X64-NEXT:    vmovsh %xmm0, (%rsi)
; X64-NEXT:    retq
  %1 = load float, ptr %val, align 8
  %res = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %1,
                                                                  metadata !"round.dynamic",
                                                                  metadata !"fpexcept.strict") #0
  store half %res, ptr %ret, align 4
  ret void
}

define void @fptrunc_double_to_f16(ptr %val, ptr%ret) nounwind strictfp {
; SSE2-LABEL: fptrunc_double_to_f16:
; SSE2:       # %bb.0:
; SSE2-NEXT:    pushq %rbx
; SSE2-NEXT:    movq %rsi, %rbx
; SSE2-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
; SSE2-NEXT:    callq __truncdfhf2@PLT
; SSE2-NEXT:    pextrw $0, %xmm0, %eax
; SSE2-NEXT:    movw %ax, (%rbx)
; SSE2-NEXT:    popq %rbx
; SSE2-NEXT:    retq
;
; AVX-LABEL: fptrunc_double_to_f16:
; AVX:       # %bb.0:
; AVX-NEXT:    pushq %rbx
; AVX-NEXT:    movq %rsi, %rbx
; AVX-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT:    callq __truncdfhf2@PLT
; AVX-NEXT:    vpextrw $0, %xmm0, (%rbx)
; AVX-NEXT:    popq %rbx
; AVX-NEXT:    retq
;
; X86-LABEL: fptrunc_double_to_f16:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
; X86-NEXT:    vcvtsd2sh %xmm0, %xmm0, %xmm0
; X86-NEXT:    vmovsh %xmm0, (%eax)
; X86-NEXT:    retl
;
; X64-LABEL: fptrunc_double_to_f16:
; X64:       # %bb.0:
; X64-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
; X64-NEXT:    vcvtsd2sh %xmm0, %xmm0, %xmm0
; X64-NEXT:    vmovsh %xmm0, (%rsi)
; X64-NEXT:    retq
  %1 = load double, ptr %val, align 8
  %res = call half @llvm.experimental.constrained.fptrunc.f16.f64(double %1,
                                                                  metadata !"round.dynamic",
                                                                  metadata !"fpexcept.strict") #0
  store half %res, ptr %ret, align 4
  ret void
}

define void @fsqrt_f16(ptr %a) nounwind strictfp {
; SSE2-LABEL: fsqrt_f16:
; SSE2:       # %bb.0:
; SSE2-NEXT:    pushq %rbx
; SSE2-NEXT:    movq %rdi, %rbx
; SSE2-NEXT:    pinsrw $0, (%rdi), %xmm0
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    sqrtss %xmm0, %xmm0
; SSE2-NEXT:    callq __truncsfhf2@PLT
; SSE2-NEXT:    pextrw $0, %xmm0, %eax
; SSE2-NEXT:    movw %ax, (%rbx)
; SSE2-NEXT:    popq %rbx
; SSE2-NEXT:    retq
;
; AVX-LABEL: fsqrt_f16:
; AVX:       # %bb.0:
; AVX-NEXT:    movzwl (%rdi), %eax
; AVX-NEXT:    vmovd %eax, %xmm0
; AVX-NEXT:    vcvtph2ps %xmm0, %xmm0
; AVX-NEXT:    vsqrtss %xmm0, %xmm0, %xmm0
; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
; AVX-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
; AVX-NEXT:    vmovd %xmm0, %eax
; AVX-NEXT:    movw %ax, (%rdi)
; AVX-NEXT:    retq
;
; X86-LABEL: fsqrt_f16:
; X86:       # %bb.0:
; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
; X86-NEXT:    vmovsh (%eax), %xmm0
; X86-NEXT:    vsqrtsh %xmm0, %xmm0, %xmm0
; X86-NEXT:    vmovsh %xmm0, (%eax)
; X86-NEXT:    retl
;
; X64-LABEL: fsqrt_f16:
; X64:       # %bb.0:
; X64-NEXT:    vmovsh (%rdi), %xmm0
; X64-NEXT:    vsqrtsh %xmm0, %xmm0, %xmm0
; X64-NEXT:    vmovsh %xmm0, (%rdi)
; X64-NEXT:    retq
  %1 = load half, ptr %a, align 4
  %res = call half @llvm.experimental.constrained.sqrt.f16(half %1,
                                                           metadata !"round.dynamic",
                                                           metadata !"fpexcept.strict") #0
  store half %res, ptr %a, align 4
  ret void
}

define half @fma_f16(half %a, half %b, half %c) nounwind strictfp {
; SSE2-LABEL: fma_f16:
; SSE2:       # %bb.0:
; SSE2-NEXT:    subq $24, %rsp
; SSE2-NEXT:    movss %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; SSE2-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; SSE2-NEXT:    movaps %xmm1, %xmm0
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
; SSE2-NEXT:    # xmm0 = mem[0],zero,zero,zero
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
; SSE2-NEXT:    # xmm0 = mem[0],zero,zero,zero
; SSE2-NEXT:    callq __extendhfsf2@PLT
; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
; SSE2-NEXT:    # xmm1 = mem[0],zero,zero,zero
; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 4-byte Reload
; SSE2-NEXT:    # xmm2 = mem[0],zero,zero,zero
; SSE2-NEXT:    callq fmaf@PLT
; SSE2-NEXT:    callq __truncsfhf2@PLT
; SSE2-NEXT:    addq $24, %rsp
; SSE2-NEXT:    retq
;
; F16C-LABEL: fma_f16:
; F16C:       # %bb.0:
; F16C-NEXT:    pushq %rax
; F16C-NEXT:    vpextrw $0, %xmm0, %eax
; F16C-NEXT:    vpextrw $0, %xmm1, %ecx
; F16C-NEXT:    vpextrw $0, %xmm2, %edx
; F16C-NEXT:    movzwl %dx, %edx
; F16C-NEXT:    vmovd %edx, %xmm0
; F16C-NEXT:    vcvtph2ps %xmm0, %xmm2
; F16C-NEXT:    movzwl %cx, %ecx
; F16C-NEXT:    vmovd %ecx, %xmm0
; F16C-NEXT:    vcvtph2ps %xmm0, %xmm1
; F16C-NEXT:    movzwl %ax, %eax
; F16C-NEXT:    vmovd %eax, %xmm0
; F16C-NEXT:    vcvtph2ps %xmm0, %xmm0
; F16C-NEXT:    callq fmaf@PLT
; F16C-NEXT:    vxorps %xmm1, %xmm1, %xmm1
; F16C-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
; F16C-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
; F16C-NEXT:    vmovd %xmm0, %eax
; F16C-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
; F16C-NEXT:    popq %rax
; F16C-NEXT:    retq
;
; AVX512-LABEL: fma_f16:
; AVX512:       # %bb.0:
; AVX512-NEXT:    vpextrw $0, %xmm1, %eax
; AVX512-NEXT:    vpextrw $0, %xmm0, %ecx
; AVX512-NEXT:    vpextrw $0, %xmm2, %edx
; AVX512-NEXT:    movzwl %dx, %edx
; AVX512-NEXT:    vmovd %edx, %xmm0
; AVX512-NEXT:    vcvtph2ps %xmm0, %xmm0
; AVX512-NEXT:    movzwl %cx, %ecx
; AVX512-NEXT:    vmovd %ecx, %xmm1
; AVX512-NEXT:    vcvtph2ps %xmm1, %xmm1
; AVX512-NEXT:    movzwl %ax, %eax
; AVX512-NEXT:    vmovd %eax, %xmm2
; AVX512-NEXT:    vcvtph2ps %xmm2, %xmm2
; AVX512-NEXT:    vfmadd213ss {{.*#+}} xmm2 = (xmm1 * xmm2) + xmm0
; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; AVX512-NEXT:    vblendps {{.*#+}} xmm0 = xmm2[0],xmm0[1,2,3]
; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
; AVX512-NEXT:    vmovd %xmm0, %eax
; AVX512-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
; AVX512-NEXT:    retq
;
; X86-LABEL: fma_f16:
; X86:       # %bb.0:
; X86-NEXT:    vmovsh {{[0-9]+}}(%esp), %xmm1
; X86-NEXT:    vmovsh {{[0-9]+}}(%esp), %xmm0
; X86-NEXT:    vfmadd213sh {{[0-9]+}}(%esp), %xmm1, %xmm0
; X86-NEXT:    retl
;
; X64-LABEL: fma_f16:
; X64:       # %bb.0:
; X64-NEXT:    vfmadd213sh %xmm2, %xmm1, %xmm0
; X64-NEXT:    retq
  %res = call half @llvm.experimental.constrained.fma.f16(half %a, half %b, half %c,
                                                          metadata !"round.dynamic",
                                                          metadata !"fpexcept.strict") #0
  ret half %res
}

attributes #0 = { strictfp }