[X86] Make zero extend from v16i1/v8i1 to v16i8/v8i16/v16i16 not scalarize under AVX512.

llvm-svn: 319136
This commit is contained in:
Craig Topper 2017-11-28 01:36:33 +00:00
parent 5befc5bfce
commit ddbc340c20
2 changed files with 38 additions and 650 deletions

View File

@ -1255,9 +1255,13 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
setOperationAction(ISD::ZERO_EXTEND, MVT::v16i8, Custom);
setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
setOperationAction(ISD::ZERO_EXTEND, MVT::v8i16, Custom);
setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
for (auto VT : { MVT::v16f32, MVT::v8f64 }) {
setOperationAction(ISD::FFLOOR, VT, Legal);

View File

@ -165,92 +165,16 @@ define <8 x i16> @ext_i8_8i16(i8 %a0) {
;
; AVX512F-LABEL: ext_i8_8i16:
; AVX512F: # BB#0:
; AVX512F-NEXT: kmovw %edi, %k5
; AVX512F-NEXT: kshiftlw $8, %k5, %k0
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kshiftlw $9, %k5, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kshiftlw $10, %k5, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kshiftlw $11, %k5, %k3
; AVX512F-NEXT: kshiftrw $15, %k3, %k3
; AVX512F-NEXT: kshiftlw $12, %k5, %k4
; AVX512F-NEXT: kshiftrw $15, %k4, %k4
; AVX512F-NEXT: kshiftlw $13, %k5, %k6
; AVX512F-NEXT: kshiftrw $15, %k6, %k6
; AVX512F-NEXT: kshiftlw $15, %k5, %k7
; AVX512F-NEXT: kshiftrw $15, %k7, %k7
; AVX512F-NEXT: kshiftlw $14, %k5, %k5
; AVX512F-NEXT: kshiftrw $15, %k5, %k5
; AVX512F-NEXT: kmovw %k5, %eax
; AVX512F-NEXT: andl $1, %eax
; AVX512F-NEXT: kmovw %k7, %ecx
; AVX512F-NEXT: andl $1, %ecx
; AVX512F-NEXT: vmovd %ecx, %xmm0
; AVX512F-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kmovw %k6, %eax
; AVX512F-NEXT: andl $1, %eax
; AVX512F-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kmovw %k4, %eax
; AVX512F-NEXT: andl $1, %eax
; AVX512F-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kmovw %k3, %eax
; AVX512F-NEXT: andl $1, %eax
; AVX512F-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: andl $1, %eax
; AVX512F-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: andl $1, %eax
; AVX512F-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kmovw %k0, %eax
; AVX512F-NEXT: andl $1, %eax
; AVX512F-NEXT: vpinsrw $7, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kmovw %edi, %k1
; AVX512F-NEXT: vpbroadcastq {{.*}}(%rip), %zmm0 {%k1} {z}
; AVX512F-NEXT: vpmovqw %zmm0, %xmm0
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VLBW-LABEL: ext_i8_8i16:
; AVX512VLBW: # BB#0:
; AVX512VLBW-NEXT: kmovd %edi, %k5
; AVX512VLBW-NEXT: kshiftlw $8, %k5, %k0
; AVX512VLBW-NEXT: kshiftrw $15, %k0, %k0
; AVX512VLBW-NEXT: kshiftlw $9, %k5, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kshiftlw $10, %k5, %k2
; AVX512VLBW-NEXT: kshiftrw $15, %k2, %k2
; AVX512VLBW-NEXT: kshiftlw $11, %k5, %k3
; AVX512VLBW-NEXT: kshiftrw $15, %k3, %k3
; AVX512VLBW-NEXT: kshiftlw $12, %k5, %k4
; AVX512VLBW-NEXT: kshiftrw $15, %k4, %k4
; AVX512VLBW-NEXT: kshiftlw $13, %k5, %k6
; AVX512VLBW-NEXT: kshiftrw $15, %k6, %k6
; AVX512VLBW-NEXT: kshiftlw $15, %k5, %k7
; AVX512VLBW-NEXT: kshiftrw $15, %k7, %k7
; AVX512VLBW-NEXT: kshiftlw $14, %k5, %k5
; AVX512VLBW-NEXT: kshiftrw $15, %k5, %k5
; AVX512VLBW-NEXT: kmovd %k5, %eax
; AVX512VLBW-NEXT: andl $1, %eax
; AVX512VLBW-NEXT: kmovd %k7, %ecx
; AVX512VLBW-NEXT: andl $1, %ecx
; AVX512VLBW-NEXT: vmovd %ecx, %xmm0
; AVX512VLBW-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0
; AVX512VLBW-NEXT: kmovd %k6, %eax
; AVX512VLBW-NEXT: andl $1, %eax
; AVX512VLBW-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0
; AVX512VLBW-NEXT: kmovd %k4, %eax
; AVX512VLBW-NEXT: andl $1, %eax
; AVX512VLBW-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0
; AVX512VLBW-NEXT: kmovd %k3, %eax
; AVX512VLBW-NEXT: andl $1, %eax
; AVX512VLBW-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0
; AVX512VLBW-NEXT: kmovd %k2, %eax
; AVX512VLBW-NEXT: andl $1, %eax
; AVX512VLBW-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0
; AVX512VLBW-NEXT: kmovd %k1, %eax
; AVX512VLBW-NEXT: andl $1, %eax
; AVX512VLBW-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
; AVX512VLBW-NEXT: kmovd %k0, %eax
; AVX512VLBW-NEXT: andl $1, %eax
; AVX512VLBW-NEXT: vpinsrw $7, %eax, %xmm0, %xmm0
; AVX512VLBW-NEXT: kmovd %edi, %k1
; AVX512VLBW-NEXT: vmovdqu16 {{.*}}(%rip), %xmm0 {%k1} {z}
; AVX512VLBW-NEXT: retq
%1 = bitcast i8 %a0 to <8 x i1>
%2 = zext <8 x i1> %1 to <8 x i16>
@ -306,188 +230,16 @@ define <16 x i8> @ext_i16_16i8(i16 %a0) {
;
; AVX512F-LABEL: ext_i16_16i8:
; AVX512F: # BB#0:
; AVX512F-NEXT: pushq %rbp
; AVX512F-NEXT: .cfi_def_cfa_offset 16
; AVX512F-NEXT: pushq %r15
; AVX512F-NEXT: .cfi_def_cfa_offset 24
; AVX512F-NEXT: pushq %r14
; AVX512F-NEXT: .cfi_def_cfa_offset 32
; AVX512F-NEXT: pushq %r13
; AVX512F-NEXT: .cfi_def_cfa_offset 40
; AVX512F-NEXT: pushq %r12
; AVX512F-NEXT: .cfi_def_cfa_offset 48
; AVX512F-NEXT: pushq %rbx
; AVX512F-NEXT: .cfi_def_cfa_offset 56
; AVX512F-NEXT: .cfi_offset %rbx, -56
; AVX512F-NEXT: .cfi_offset %r12, -48
; AVX512F-NEXT: .cfi_offset %r13, -40
; AVX512F-NEXT: .cfi_offset %r14, -32
; AVX512F-NEXT: .cfi_offset %r15, -24
; AVX512F-NEXT: .cfi_offset %rbp, -16
; AVX512F-NEXT: kmovw %edi, %k0
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %r8d
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %r9d
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %r10d
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %r11d
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %r14d
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %r15d
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %r12d
; AVX512F-NEXT: kshiftlw $8, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %r13d
; AVX512F-NEXT: kshiftlw $7, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %esi
; AVX512F-NEXT: kshiftlw $6, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %ebx
; AVX512F-NEXT: kshiftlw $5, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %ebp
; AVX512F-NEXT: kshiftlw $4, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %edi
; AVX512F-NEXT: kshiftlw $3, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: kshiftlw $2, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %ecx
; AVX512F-NEXT: kshiftlw $1, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %edx
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: vmovd %r9d, %xmm0
; AVX512F-NEXT: kmovw %k0, %r9d
; AVX512F-NEXT: vpinsrb $1, %r8d, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $2, %r10d, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $3, %r11d, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $4, %r14d, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $5, %r15d, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $6, %r12d, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $7, %r13d, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $8, %esi, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $9, %ebx, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $10, %ebp, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $11, %edi, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $12, %eax, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $13, %ecx, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $14, %edx, %xmm0, %xmm0
; AVX512F-NEXT: vpinsrb $15, %r9d, %xmm0, %xmm0
; AVX512F-NEXT: vpand {{.*}}(%rip), %xmm0, %xmm0
; AVX512F-NEXT: popq %rbx
; AVX512F-NEXT: popq %r12
; AVX512F-NEXT: popq %r13
; AVX512F-NEXT: popq %r14
; AVX512F-NEXT: popq %r15
; AVX512F-NEXT: popq %rbp
; AVX512F-NEXT: kmovw %edi, %k1
; AVX512F-NEXT: vpbroadcastd {{.*}}(%rip), %zmm0 {%k1} {z}
; AVX512F-NEXT: vpmovdb %zmm0, %xmm0
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512VLBW-LABEL: ext_i16_16i8:
; AVX512VLBW: # BB#0:
; AVX512VLBW-NEXT: pushq %rbp
; AVX512VLBW-NEXT: .cfi_def_cfa_offset 16
; AVX512VLBW-NEXT: pushq %r15
; AVX512VLBW-NEXT: .cfi_def_cfa_offset 24
; AVX512VLBW-NEXT: pushq %r14
; AVX512VLBW-NEXT: .cfi_def_cfa_offset 32
; AVX512VLBW-NEXT: pushq %r13
; AVX512VLBW-NEXT: .cfi_def_cfa_offset 40
; AVX512VLBW-NEXT: pushq %r12
; AVX512VLBW-NEXT: .cfi_def_cfa_offset 48
; AVX512VLBW-NEXT: pushq %rbx
; AVX512VLBW-NEXT: .cfi_def_cfa_offset 56
; AVX512VLBW-NEXT: .cfi_offset %rbx, -56
; AVX512VLBW-NEXT: .cfi_offset %r12, -48
; AVX512VLBW-NEXT: .cfi_offset %r13, -40
; AVX512VLBW-NEXT: .cfi_offset %r14, -32
; AVX512VLBW-NEXT: .cfi_offset %r15, -24
; AVX512VLBW-NEXT: .cfi_offset %rbp, -16
; AVX512VLBW-NEXT: kmovd %edi, %k0
; AVX512VLBW-NEXT: kshiftlw $14, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %r8d
; AVX512VLBW-NEXT: kshiftlw $15, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %r9d
; AVX512VLBW-NEXT: kshiftlw $13, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %r10d
; AVX512VLBW-NEXT: kshiftlw $12, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %r11d
; AVX512VLBW-NEXT: kshiftlw $11, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %r14d
; AVX512VLBW-NEXT: kshiftlw $10, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %r15d
; AVX512VLBW-NEXT: kshiftlw $9, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %r12d
; AVX512VLBW-NEXT: kshiftlw $8, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %r13d
; AVX512VLBW-NEXT: kshiftlw $7, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %esi
; AVX512VLBW-NEXT: kshiftlw $6, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %ebx
; AVX512VLBW-NEXT: kshiftlw $5, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %ebp
; AVX512VLBW-NEXT: kshiftlw $4, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %edi
; AVX512VLBW-NEXT: kshiftlw $3, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %eax
; AVX512VLBW-NEXT: kshiftlw $2, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %ecx
; AVX512VLBW-NEXT: kshiftlw $1, %k0, %k1
; AVX512VLBW-NEXT: kshiftrw $15, %k1, %k1
; AVX512VLBW-NEXT: kmovd %k1, %edx
; AVX512VLBW-NEXT: kshiftrw $15, %k0, %k0
; AVX512VLBW-NEXT: vmovd %r9d, %xmm0
; AVX512VLBW-NEXT: kmovd %k0, %r9d
; AVX512VLBW-NEXT: vpinsrb $1, %r8d, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $2, %r10d, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $3, %r11d, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $4, %r14d, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $5, %r15d, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $6, %r12d, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $7, %r13d, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $8, %esi, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $9, %ebx, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $10, %ebp, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $11, %edi, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $12, %eax, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $13, %ecx, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $14, %edx, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpinsrb $15, %r9d, %xmm0, %xmm0
; AVX512VLBW-NEXT: vpand {{.*}}(%rip), %xmm0, %xmm0
; AVX512VLBW-NEXT: popq %rbx
; AVX512VLBW-NEXT: popq %r12
; AVX512VLBW-NEXT: popq %r13
; AVX512VLBW-NEXT: popq %r14
; AVX512VLBW-NEXT: popq %r15
; AVX512VLBW-NEXT: popq %rbp
; AVX512VLBW-NEXT: kmovd %edi, %k1
; AVX512VLBW-NEXT: vmovdqu8 {{.*}}(%rip), %xmm0 {%k1} {z}
; AVX512VLBW-NEXT: retq
%1 = bitcast i16 %a0 to <16 x i1>
%2 = zext <16 x i1> %1 to <16 x i8>
@ -769,136 +521,14 @@ define <32 x i8> @ext_i32_32i8(i32 %a0) {
; AVX512F-NEXT: andq $-32, %rsp
; AVX512F-NEXT: subq $32, %rsp
; AVX512F-NEXT: movl %edi, (%rsp)
; AVX512F-NEXT: kmovw (%rsp), %k0
; AVX512F-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; AVX512F-NEXT: kshiftlw $14, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: kshiftlw $15, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %ecx
; AVX512F-NEXT: vmovd %ecx, %xmm0
; AVX512F-NEXT: vpinsrb $1, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $13, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $2, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $12, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $3, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $11, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $4, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $10, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $5, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $9, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $6, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $8, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $7, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $7, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $8, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $6, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $9, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $5, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $10, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $4, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $11, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $3, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $12, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $2, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $13, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $1, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $14, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %ecx
; AVX512F-NEXT: vmovd %ecx, %xmm1
; AVX512F-NEXT: vpinsrb $1, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $2, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $3, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $5, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $6, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $8, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $7, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $7, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $6, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $9, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $5, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $10, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $4, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $11, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $3, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $2, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $13, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $1, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $14, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, %eax
; AVX512F-NEXT: vpinsrb $15, %eax, %xmm1, %xmm1
; AVX512F-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX512F-NEXT: vpand {{.*}}(%rip), %ymm0, %ymm0
; AVX512F-NEXT: movl {{.*}}(%rip), %eax
; AVX512F-NEXT: kmovw (%rsp), %k1
; AVX512F-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; AVX512F-NEXT: vpbroadcastd %eax, %zmm0 {%k1} {z}
; AVX512F-NEXT: vpmovdb %zmm0, %xmm0
; AVX512F-NEXT: vpbroadcastd %eax, %zmm1 {%k2} {z}
; AVX512F-NEXT: vpmovdb %zmm1, %xmm1
; AVX512F-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512F-NEXT: movq %rbp, %rsp
; AVX512F-NEXT: popq %rbp
; AVX512F-NEXT: retq
@ -1316,267 +946,21 @@ define <64 x i8> @ext_i64_64i8(i64 %a0) {
; AVX512F-NEXT: movl %edi, (%rsp)
; AVX512F-NEXT: shrq $32, %rdi
; AVX512F-NEXT: movl %edi, {{[0-9]+}}(%rsp)
; AVX512F-NEXT: kmovw (%rsp), %k0
; AVX512F-NEXT: movl {{.*}}(%rip), %eax
; AVX512F-NEXT: kmovw (%rsp), %k1
; AVX512F-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; AVX512F-NEXT: vpbroadcastd %eax, %zmm0 {%k1} {z}
; AVX512F-NEXT: vpmovdb %zmm0, %xmm0
; AVX512F-NEXT: vpbroadcastd %eax, %zmm1 {%k2} {z}
; AVX512F-NEXT: vpmovdb %zmm1, %xmm1
; AVX512F-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512F-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; AVX512F-NEXT: kshiftlw $14, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: kshiftlw $15, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %ecx
; AVX512F-NEXT: vmovd %ecx, %xmm0
; AVX512F-NEXT: vpinsrb $1, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $13, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $2, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $12, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $3, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $11, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $4, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $10, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $5, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $9, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $6, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $8, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $7, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $7, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $8, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $6, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $9, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $5, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $10, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $4, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $11, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $3, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $12, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $2, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $13, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $1, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $14, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %ecx
; AVX512F-NEXT: vmovd %ecx, %xmm1
; AVX512F-NEXT: vpinsrb $1, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $2, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $3, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $5, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $6, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $8, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $7, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $7, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $6, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $9, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $5, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $10, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $4, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $11, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $3, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $2, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $13, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftlw $1, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $14, %eax, %xmm1, %xmm1
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, %eax
; AVX512F-NEXT: vpinsrb $15, %eax, %xmm1, %xmm1
; AVX512F-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
; AVX512F-NEXT: vpand %ymm1, %ymm0, %ymm0
; AVX512F-NEXT: kmovw {{[0-9]+}}(%rsp), %k0
; AVX512F-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; AVX512F-NEXT: kshiftlw $14, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: kshiftlw $15, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %ecx
; AVX512F-NEXT: vmovd %ecx, %xmm2
; AVX512F-NEXT: vpinsrb $1, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $13, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $2, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $12, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $3, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $11, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $4, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $10, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $5, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $9, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $6, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $8, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $7, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $7, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $8, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $6, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $9, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $5, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $10, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $4, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $11, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $3, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $12, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $2, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $1, %k1, %k2
; AVX512F-NEXT: kshiftrw $15, %k2, %k2
; AVX512F-NEXT: kmovw %k2, %eax
; AVX512F-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2
; AVX512F-NEXT: kshiftlw $14, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: kshiftlw $15, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %ecx
; AVX512F-NEXT: vmovd %ecx, %xmm3
; AVX512F-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $13, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $12, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $11, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $10, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $9, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $8, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $7, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $6, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $5, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $4, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $3, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $2, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftlw $1, %k0, %k1
; AVX512F-NEXT: kshiftrw $15, %k1, %k1
; AVX512F-NEXT: kmovw %k1, %eax
; AVX512F-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3
; AVX512F-NEXT: kshiftrw $15, %k0, %k0
; AVX512F-NEXT: kmovw %k0, %eax
; AVX512F-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
; AVX512F-NEXT: vpand %ymm1, %ymm2, %ymm1
; AVX512F-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; AVX512F-NEXT: vpbroadcastd %eax, %zmm1 {%k1} {z}
; AVX512F-NEXT: vpmovdb %zmm1, %xmm1
; AVX512F-NEXT: vpbroadcastd %eax, %zmm2 {%k2} {z}
; AVX512F-NEXT: vpmovdb %zmm2, %xmm2
; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: movq %rbp, %rsp
; AVX512F-NEXT: popq %rbp
; AVX512F-NEXT: retq