[X86][FP16] Add the missing legal action for EXTRACT_SUBVECTOR

[pengfei]

Summary

Fixes #57340

[RKSimon]

LGTM

[pengfei]

Upload too quick to catch the other updates :)

[pengfei]

Thanks @RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64,
                      MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64 }) {
       setOperationAction(ISD::MLOAD,  VT, Subtarget.hasVLX() ? Legal : Custom);
       setOperationAction(ISD::MSTORE, VT, Legal);
     }
 
     // Extract subvector is special because the value type
     // (result) is 128-bit but the source is 256-bit wide.
     for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64,
-                     MVT::v4f32, MVT::v2f64 }) {
+                     MVT::v8f16, MVT::v4f32, MVT::v2f64 }) {
       setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
     }
 
     // Custom lower several nodes for 256-bit types.
     for (MVT VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64,
                     MVT::v16f16, MVT::v8f32, MVT::v4f64 }) {
       setOperationAction(ISD::BUILD_VECTOR,       VT, Custom);
       setOperationAction(ISD::VECTOR_SHUFFLE,     VT, Custom);
if (Subtarget.hasVPOPCNTDQ()) {
       for (auto VT : { MVT::v16i32, MVT::v8i64 })
         setOperationAction(ISD::CTPOP, VT, Legal);
     }
 
     // Extract subvector is special because the value type
     // (result) is 256-bit but the source is 512-bit wide.
     // 128-bit was made Legal under AVX1.
     for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64,
-                     MVT::v8f32, MVT::v4f64 })
+                     MVT::v16f16, MVT::v8f32, MVT::v4f64 })
       setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
 
     for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32, MVT::v8i64,
                      MVT::v32f16, MVT::v16f32, MVT::v8f64 }) {
       setOperationAction(ISD::CONCAT_VECTORS,     VT, Custom);
       setOperationAction(ISD::INSERT_SUBVECTOR,   VT, Legal);
       setOperationAction(ISD::SELECT,             VT, Custom);
       setOperationAction(ISD::VSELECT,            VT, Custom);

llvm/test/CodeGen/X86/avx512-f16c-v16f16-fadd.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-- -mattr=f16c| FileCheck %s --check-prefixes=CHECK
 
 define <16 x half> @foo(<16 x half> %a, <16 x half> %b) nounwind {
 ; CHECK-LABEL: foo:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    pushq %rbp
-; CHECK-NEXT:    movq %rsp, %rbp
-; CHECK-NEXT:    andq $-32, %rsp
-; CHECK-NEXT:    subq $96, %rsp
-; CHECK-NEXT:    vmovaps %ymm1, {{[0-9]+}}(%rsp)
-; CHECK-NEXT:    vmovaps %ymm0, (%rsp)
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %ymm0
-; CHECK-NEXT:    vcvtph2ps (%rsp), %ymm1
-; CHECK-NEXT:    vaddps %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    vcvtph2ps %xmm1, %ymm2
+; CHECK-NEXT:    vcvtph2ps %xmm0, %ymm3
+; CHECK-NEXT:    vaddps %ymm2, %ymm3, %ymm2
+; CHECK-NEXT:    vcvtps2ph $4, %ymm2, %xmm2
+; CHECK-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; CHECK-NEXT:    vcvtph2ps %xmm1, %ymm1
+; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; CHECK-NEXT:    vcvtph2ps %xmm0, %ymm0
+; CHECK-NEXT:    vaddps %ymm1, %ymm0, %ymm0
 ; CHECK-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %ymm1
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %ymm2
-; CHECK-NEXT:    vaddps %ymm1, %ymm2, %ymm1
-; CHECK-NEXT:    vcvtps2ph $4, %ymm1, %xmm1
-; CHECK-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; CHECK-NEXT:    movq %rbp, %rsp
-; CHECK-NEXT:    popq %rbp
+; CHECK-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
 ; CHECK-NEXT:    retq
   %1 = fadd <16 x half> %a, %b
   ret <16 x half> %1
 }

llvm/test/CodeGen/X86/avx512-skx-v32f16-fadd.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-- -mattr=avx512f| FileCheck %s --check-prefixes=CHECK
 
 define <32 x half> @foo(<32 x half> %a, <32 x half> %b) nounwind {
 ; CHECK-LABEL: foo:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    pushq %rbp
-; CHECK-NEXT:    movq %rsp, %rbp
-; CHECK-NEXT:    andq $-64, %rsp
-; CHECK-NEXT:    subq $192, %rsp
-; CHECK-NEXT:    vmovaps %zmm1, {{[0-9]+}}(%rsp)
-; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %zmm0
-; CHECK-NEXT:    vcvtph2ps (%rsp), %zmm1
-; CHECK-NEXT:    vaddps %zmm0, %zmm1, %zmm0
+; CHECK-NEXT:    vcvtph2ps %ymm1, %zmm2
+; CHECK-NEXT:    vcvtph2ps %ymm0, %zmm3
+; CHECK-NEXT:    vaddps %zmm2, %zmm3, %zmm2
+; CHECK-NEXT:    vcvtps2ph $4, %zmm2, %ymm2
+; CHECK-NEXT:    vextractf64x4 $1, %zmm1, %ymm1
+; CHECK-NEXT:    vcvtph2ps %ymm1, %zmm1
+; CHECK-NEXT:    vextractf64x4 $1, %zmm0, %ymm0
+; CHECK-NEXT:    vcvtph2ps %ymm0, %zmm0
+; CHECK-NEXT:    vaddps %zmm1, %zmm0, %zmm0
 ; CHECK-NEXT:    vcvtps2ph $4, %zmm0, %ymm0
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %zmm1
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %zmm2
-; CHECK-NEXT:    vaddps %zmm1, %zmm2, %zmm1
-; CHECK-NEXT:    vcvtps2ph $4, %zmm1, %ymm1
-; CHECK-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
-; CHECK-NEXT:    movq %rbp, %rsp
-; CHECK-NEXT:    popq %rbp
+; CHECK-NEXT:    vinsertf64x4 $1, %ymm0, %zmm2, %zmm0
 ; CHECK-NEXT:    retq
   %1 = fadd <32 x half> %a, %b
   ret <32 x half> %1
 }

llvm/test/CodeGen/X86/pr57340.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-unknown | FileCheck %s
+
+define void @main.41() local_unnamed_addr #1 {
+; CHECK-LABEL: main.41:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vpbroadcastw (%rax), %xmm0
+; CHECK-NEXT:    vpextrw $0, %xmm0, %eax
+; CHECK-NEXT:    vmovdqu (%rax), %ymm2
+; CHECK-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
+; CHECK-NEXT:    vmovdqa {{.*#+}} ymm1 = [31,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14]
+; CHECK-NEXT:    vpermi2w %ymm0, %ymm2, %ymm1
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm0
+; CHECK-NEXT:    vcvtph2ps %xmm0, %xmm0
+; CHECK-NEXT:    vmovdqu (%rax), %xmm6
+; CHECK-NEXT:    vpextrw $0, %xmm6, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm2
+; CHECK-NEXT:    vcvtph2ps %xmm2, %xmm2
+; CHECK-NEXT:    vucomiss %xmm0, %xmm2
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    andl $1, %eax
+; CHECK-NEXT:    kmovw %eax, %k0
+; CHECK-NEXT:    vpsrld $16, %xmm1, %xmm0
+; CHECK-NEXT:    vpextrw $0, %xmm0, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm0
+; CHECK-NEXT:    vcvtph2ps %xmm0, %xmm0
+; CHECK-NEXT:    vpsrld $16, %xmm6, %xmm3
+; CHECK-NEXT:    vpextrw $0, %xmm3, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm3
+; CHECK-NEXT:    vcvtph2ps %xmm3, %xmm3
+; CHECK-NEXT:    vucomiss %xmm0, %xmm3
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $14, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-5, %ax
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    vpshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
+; CHECK-NEXT:    vpextrw $0, %xmm0, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm0
+; CHECK-NEXT:    vcvtph2ps %xmm0, %xmm4
+; CHECK-NEXT:    movzwl (%rax), %eax
+; CHECK-NEXT:    vmovd %eax, %xmm0
+; CHECK-NEXT:    vcvtph2ps %xmm0, %xmm0
+; CHECK-NEXT:    vucomiss %xmm4, %xmm0
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $13, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-9, %ax
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    vpsrlq $48, %xmm1, %xmm4
+; CHECK-NEXT:    vpextrw $0, %xmm4, %eax
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm4
+; CHECK-NEXT:    vcvtph2ps %xmm4, %xmm5
+; CHECK-NEXT:    vpsrlq $48, %xmm6, %xmm4
+; CHECK-NEXT:    vpextrw $0, %xmm4, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm4
+; CHECK-NEXT:    vcvtph2ps %xmm4, %xmm4
+; CHECK-NEXT:    vucomiss %xmm5, %xmm4
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $12, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-17, %ax
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    vpshufd {{.*#+}} xmm5 = xmm1[2,3,0,1]
+; CHECK-NEXT:    vpextrw $0, %xmm5, %eax
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm5
+; CHECK-NEXT:    vcvtph2ps %xmm5, %xmm5
+; CHECK-NEXT:    vucomiss %xmm5, %xmm0
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $11, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-33, %ax
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; CHECK-NEXT:    vpextrw $0, %xmm5, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm5
+; CHECK-NEXT:    vcvtph2ps %xmm5, %xmm7
+; CHECK-NEXT:    vpsrldq {{.*#+}} xmm5 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; CHECK-NEXT:    vpextrw $0, %xmm5, %eax
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm5
+; CHECK-NEXT:    vcvtph2ps %xmm5, %xmm5
+; CHECK-NEXT:    vucomiss %xmm7, %xmm5
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $10, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-65, %ax
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    vpshufd {{.*#+}} xmm7 = xmm1[3,3,3,3]
+; CHECK-NEXT:    vpextrw $0, %xmm7, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm7
+; CHECK-NEXT:    vcvtph2ps %xmm7, %xmm7
+; CHECK-NEXT:    vucomiss %xmm7, %xmm0
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $9, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-129, %ax
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    vpsrldq {{.*#+}} xmm7 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; CHECK-NEXT:    vpextrw $0, %xmm7, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm7
+; CHECK-NEXT:    vcvtph2ps %xmm7, %xmm7
+; CHECK-NEXT:    vpsrldq {{.*#+}} xmm6 = xmm6[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; CHECK-NEXT:    vpextrw $0, %xmm6, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm6
+; CHECK-NEXT:    vcvtph2ps %xmm6, %xmm6
+; CHECK-NEXT:    vucomiss %xmm7, %xmm6
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $8, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-257, %ax # imm = 0xFEFF
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    vextracti128 $1, %ymm1, %xmm1
+; CHECK-NEXT:    vpextrw $0, %xmm1, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm7
+; CHECK-NEXT:    vcvtph2ps %xmm7, %xmm7
+; CHECK-NEXT:    vucomiss %xmm7, %xmm2
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $7, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-513, %ax # imm = 0xFDFF
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    vpsrld $16, %xmm1, %xmm2
+; CHECK-NEXT:    vpextrw $0, %xmm2, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm2
+; CHECK-NEXT:    vcvtph2ps %xmm2, %xmm2
+; CHECK-NEXT:    vucomiss %xmm2, %xmm3
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $6, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-1025, %ax # imm = 0xFBFF
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    vpshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
+; CHECK-NEXT:    vpextrw $0, %xmm2, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm2
+; CHECK-NEXT:    vcvtph2ps %xmm2, %xmm2
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    vucomiss %xmm2, %xmm0
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $5, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-2049, %ax # imm = 0xF7FF
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    vpsrlq $48, %xmm1, %xmm2
+; CHECK-NEXT:    vpextrw $0, %xmm2, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm2
+; CHECK-NEXT:    vcvtph2ps %xmm2, %xmm2
+; CHECK-NEXT:    vucomiss %xmm2, %xmm4
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $4, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-4097, %ax # imm = 0xEFFF
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    vpshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
+; CHECK-NEXT:    vpextrw $0, %xmm2, %eax
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm2
+; CHECK-NEXT:    vcvtph2ps %xmm2, %xmm2
+; CHECK-NEXT:    vucomiss %xmm2, %xmm0
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $3, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-8193, %ax # imm = 0xDFFF
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    vpsrldq {{.*#+}} xmm2 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; CHECK-NEXT:    vpextrw $0, %xmm2, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm2
+; CHECK-NEXT:    vcvtph2ps %xmm2, %xmm2
+; CHECK-NEXT:    vucomiss %xmm2, %xmm5
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    kshiftrw $2, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    movw $-16385, %ax # imm = 0xBFFF
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    vpshufd {{.*#+}} xmm2 = xmm1[3,3,3,3]
+; CHECK-NEXT:    vpextrw $0, %xmm2, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm2
+; CHECK-NEXT:    vcvtph2ps %xmm2, %xmm2
+; CHECK-NEXT:    vucomiss %xmm2, %xmm0
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $14, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k0
+; CHECK-NEXT:    kshiftlw $1, %k0, %k0
+; CHECK-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; CHECK-NEXT:    vpextrw $0, %xmm0, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm0
+; CHECK-NEXT:    vcvtph2ps %xmm0, %xmm0
+; CHECK-NEXT:    kshiftrw $1, %k0, %k0
+; CHECK-NEXT:    vucomiss %xmm0, %xmm6
+; CHECK-NEXT:    setnp %al
+; CHECK-NEXT:    sete %cl
+; CHECK-NEXT:    testb %al, %cl
+; CHECK-NEXT:    setne %al
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    kshiftlw $15, %k1, %k1
+; CHECK-NEXT:    korw %k1, %k0, %k1
+; CHECK-NEXT:    vmovdqu8 {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 {%k1} {z}
+; CHECK-NEXT:    vmovdqa %xmm0, (%rax)
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+entry:
+  %.pre = load half, ptr undef, align 16
+  %vector.recur.init = insertelement <16 x half> poison, half %.pre, i64 15
+  %wide.load = load <16 x half>, ptr undef, align 2
+  %0 = shufflevector <16 x half> %vector.recur.init, <16 x half> %wide.load, <16 x i32> <i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30>
+  %1 = fcmp oeq <16 x half> %wide.load, %0
+  %2 = zext <16 x i1> %1 to <16 x i8>
+  store <16 x i8> %2, ptr undef, align 16
+  ret void
+}
+
+attributes #1 = { nounwind uwtable "target-cpu"="skx" }