test/CodeGen/X86/avx512vl-vec-masked-cmp.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -disable-peephole -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq | FileCheck %s --check-prefix=CHECK --check-prefix=VLX
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -disable-peephole -mattr=+avx512f | FileCheck %s --check-prefix=CHECK --check-prefix=NoVLX

define zeroext i32 @test_vpcmpeqb_v16i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqb_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqb %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqb_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp eq <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpeqb_v16i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqb_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqb (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqb_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpeqb (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp eq <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqb_v16i1_v32i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqb_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqb %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqb_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp eq <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpeqb_v16i1_v32i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqb_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqb (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqb_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqb (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp eq <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpeqb_v16i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqb_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqb %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqb_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp eq <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqb_v16i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqb_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqb (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqb_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpeqb (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp eq <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqb_v16i1_v64i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqb_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqb %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqb_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp eq <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqb_v16i1_v64i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqb_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqb (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqb_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqb (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp eq <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpeqb_v32i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqb_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqb %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqb_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %1 = bitcast <4 x i64> %__b to <32 x i8>
  %2 = icmp eq <32 x i8> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqb_v32i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqb_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqb (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqb_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqb (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <32 x i8>
  %2 = icmp eq <32 x i8> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqb_v32i1_v64i1_mask(i32 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqb_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqb %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqb_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm3, %xmm3
; NoVLX-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %1 = bitcast <4 x i64> %__b to <32 x i8>
  %2 = icmp eq <32 x i8> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqb_v32i1_v64i1_mask_mem(i32 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqb_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqb (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqb_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    vpcmpeqb (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vpand %xmm2, %xmm3, %xmm2
; NoVLX-NEXT:    vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT:    vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT:    vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <32 x i8>
  %2 = icmp eq <32 x i8> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpeqw_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpeqw_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqw (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqw_v8i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpeqw_v8i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqw (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpeqw_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpeqw_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqw (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqw_v8i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpeqw_v8i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqw (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpeqw_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqw_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqw (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqw_v8i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqw_v8i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqw (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp eq <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpeqw_v16i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp eq <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpeqw_v16i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpeqw (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp eq <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqw_v16i1_v32i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp eq <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpeqw_v16i1_v32i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqw (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp eq <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpeqw_v16i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp eq <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqw_v16i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpeqw (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp eq <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqw_v16i1_v64i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp eq <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqw_v16i1_v64i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqw (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp eq <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpeqw_v32i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vmovq %xmm3, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm2
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm5
; NoVLX-NEXT:    vextracti32x4 $2, %zmm1, %xmm8
; NoVLX-NEXT:    vextracti32x4 $3, %zmm1, %xmm4
; NoVLX-NEXT:    vextracti128 $1, %ymm1, %xmm6
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm7
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm2
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm2, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm7, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm5
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm7, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm6, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm7
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm6, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm6
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm8, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm8, %rax
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
; NoVLX-NEXT:    vinserti128 $1, %xmm5, %ymm7, %ymm3
; NoVLX-NEXT:    vinserti128 $1, %xmm6, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpeqw %ymm1, %ymm3, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %1 = bitcast <8 x i64> %__b to <32 x i16>
  %2 = icmp eq <32 x i16> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqw_v32i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqw_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqw (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqw_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm2
; NoVLX-NEXT:    vmovq %xmm2, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm1
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm1, %xmm3
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm1
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm4
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpeqw 32(%rdi), %ymm1, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpeqw (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <32 x i16>
  %2 = icmp eq <32 x i16> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqw_v32i1_v64i1_mask(i32 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm2
; NoVLX-NEXT:    vmovq %xmm2, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm3
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm4
; NoVLX-NEXT:    vextracti128 $1, %ymm1, %xmm8
; NoVLX-NEXT:    vextracti32x4 $2, %zmm1, %xmm5
; NoVLX-NEXT:    vextracti32x4 $3, %zmm1, %xmm7
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm6
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm3, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm9
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm6, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm4
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm6, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm6
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm7, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm7, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm5, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm7
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm5, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm8, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm8, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    vinserti128 $1, %xmm9, %ymm4, %ymm1
; NoVLX-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm0
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vinserti128 $1, %xmm7, %ymm3, %ymm3
; NoVLX-NEXT:    vpternlogd $255, %zmm4, %zmm4, %zmm4 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm4, %xmm4
; NoVLX-NEXT:    vpcmpeqw %ymm3, %ymm1, %ymm1
; NoVLX-NEXT:    vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm3, %xmm3
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vinserti128 $1, %xmm5, %ymm2, %ymm2
; NoVLX-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpand %xmm4, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %1 = bitcast <8 x i64> %__b to <32 x i16>
  %2 = icmp eq <32 x i16> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqw_v32i1_v64i1_mask_mem(i32 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqw_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqw (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqw_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
; NoVLX-NEXT:    vmovq %xmm1, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm2
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm4
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm3, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm4, %zmm4, %zmm4 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm4, %xmm4
; NoVLX-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
; NoVLX-NEXT:    vpcmpeqw (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpand %xmm4, %xmm0, %xmm0
; NoVLX-NEXT:    vpcmpeqw 32(%rsi), %ymm1, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <32 x i16>
  %2 = icmp eq <32 x i16> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i8 @test_vpcmpeqd_v4i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpeqd_v4i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpeqd_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpeqd_v4i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpeqd_v4i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqd_v4i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpeqd_v4i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpeqd_v4i1_v16i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqd_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpeqd_v4i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpeqd_v4i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqd_v4i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpeqd_v4i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpeqd_v4i1_v32i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqd_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpeqd_v4i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqd_v4i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqd_v4i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqd_v4i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpeqd_v4i1_v64i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqd_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpeqd_v8i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpeqd_v8i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqd_v8i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpeqd_v8i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpeqd_v8i1_v16i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqd_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpeqd_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpeqd_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqd_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpeqd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpeqd_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpeqd_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqd_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqd_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpeqd_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpeqd_v16i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpeqd_v16i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqd (%rdi), %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqd_v16i1_v32i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpeqd_v16i1_v32i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd (%rsi), %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpeqd_v16i1_v32i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi){1to16}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqd (%rdi){1to16}, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqd_v16i1_v32i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd (%rsi){1to16}, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %3, %2
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpeqd_v16i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqd_v16i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqd (%rdi), %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqd_v16i1_v64i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqd_v16i1_v64i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd (%rsi), %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpeqd_v16i1_v64i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqd (%rdi){1to16}, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqd (%rdi){1to16}, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqd_v16i1_v64i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqd (%rsi){1to16}, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %3, %2
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i4 @test_vpcmpeqq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_vpcmpeqq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_masked_vpcmpeqq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}

define zeroext i4 @test_masked_vpcmpeqq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}


define zeroext i4 @test_vpcmpeqq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_masked_vpcmpeqq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}


define zeroext i8 @test_vpcmpeqq_v2i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpeqq_v2i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpeqq_v2i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpeqq_v2i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpeqq_v2i1_v8i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpeqq_v2i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpeqq_v2i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpeqq_v2i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqq_v2i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpeqq_v2i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpeqq_v2i1_v16i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqq_v2i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpeqq_v2i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpeqq_v2i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqq_v2i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpeqq_v2i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpeqq_v2i1_v32i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqq_v2i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpeqq_v2i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqq_v2i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqq_v2i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqq_v2i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpeqq_v2i1_v64i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqq_v2i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp eq <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i8 @test_vpcmpeqq_v4i1_v8i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpeqq_v4i1_v8i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpeqq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpeqq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpeqq_v4i1_v8i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpeqq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpeqq_v4i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpeqq_v4i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpeqq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpeqq_v4i1_v16i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpeqq_v4i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpeqq_v4i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpeqq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpeqq_v4i1_v32i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpeqq_v4i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqq_v4i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpeqq_v4i1_v64i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpeqq_v8i1_v16i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpeqq_v8i1_v16i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqq_v8i1_v16i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpeqq_v8i1_v16i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpeqq_v8i1_v16i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpeqq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpeqq_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpeqq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpeqq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpeqq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpeqq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpeqq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpeqq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpeqq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpeqq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpeqq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpeqq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpeqq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpeqq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp eq <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpsgtb_v16i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtb_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtb %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtb_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp sgt <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgtb_v16i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtb_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtb (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtb_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtb (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp sgt <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtb_v16i1_v32i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtb_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtb %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtb_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp sgt <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgtb_v16i1_v32i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtb_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtb (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtb_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtb (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp sgt <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgtb_v16i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtb_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtb %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtb_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp sgt <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtb_v16i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtb_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtb (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtb_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtb (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp sgt <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtb_v16i1_v64i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtb_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtb %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtb_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp sgt <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtb_v16i1_v64i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtb_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtb (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtb_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtb (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp sgt <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgtb_v32i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtb_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtb %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtb_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtb %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %1 = bitcast <4 x i64> %__b to <32 x i8>
  %2 = icmp sgt <32 x i8> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtb_v32i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtb_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtb (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtb_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtb (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <32 x i8>
  %2 = icmp sgt <32 x i8> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtb_v32i1_v64i1_mask(i32 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtb_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtb %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtb_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm3, %xmm3
; NoVLX-NEXT:    vpcmpgtb %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %1 = bitcast <4 x i64> %__b to <32 x i8>
  %2 = icmp sgt <32 x i8> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtb_v32i1_v64i1_mask_mem(i32 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtb_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtb (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtb_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    vpcmpgtb (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vpand %xmm2, %xmm3, %xmm2
; NoVLX-NEXT:    vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT:    vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT:    vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <32 x i8>
  %2 = icmp sgt <32 x i8> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpsgtw_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsgtw_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtw (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtw_v8i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsgtw_v8i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtw (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsgtw_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgtw_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtw (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtw_v8i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgtw_v8i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtw (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgtw_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtw_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtw (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtw_v8i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtw_v8i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtw (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sgt <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpsgtw_v16i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp sgt <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgtw_v16i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtw (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp sgt <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtw_v16i1_v32i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp sgt <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgtw_v16i1_v32i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtw (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp sgt <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgtw_v16i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp sgt <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtw_v16i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtw (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp sgt <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtw_v16i1_v64i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp sgt <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtw_v16i1_v64i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtw (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp sgt <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgtw_v32i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vmovq %xmm3, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm2
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm5
; NoVLX-NEXT:    vextracti32x4 $2, %zmm1, %xmm8
; NoVLX-NEXT:    vextracti32x4 $3, %zmm1, %xmm4
; NoVLX-NEXT:    vextracti128 $1, %ymm1, %xmm6
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm7
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm2
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm2, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm7, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm5
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm7, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm6, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm7
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm6, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm6
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm8, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm8, %rax
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
; NoVLX-NEXT:    vinserti128 $1, %xmm5, %ymm7, %ymm3
; NoVLX-NEXT:    vinserti128 $1, %xmm6, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm3, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %1 = bitcast <8 x i64> %__b to <32 x i16>
  %2 = icmp sgt <32 x i16> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtw_v32i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtw_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtw (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtw_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm2
; NoVLX-NEXT:    vmovq %xmm2, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm1
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm1, %xmm3
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm1
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm4
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw 32(%rdi), %ymm1, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtw (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <32 x i16>
  %2 = icmp sgt <32 x i16> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtw_v32i1_v64i1_mask(i32 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm2
; NoVLX-NEXT:    vmovq %xmm2, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm3
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm4
; NoVLX-NEXT:    vextracti128 $1, %ymm1, %xmm8
; NoVLX-NEXT:    vextracti32x4 $2, %zmm1, %xmm5
; NoVLX-NEXT:    vextracti32x4 $3, %zmm1, %xmm7
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm6
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm3, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm9
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm6, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm4
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm6, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm6
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm7, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm7, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm5, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm7
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm5, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm8, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm8, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    vinserti128 $1, %xmm9, %ymm4, %ymm1
; NoVLX-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm0
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vinserti128 $1, %xmm7, %ymm3, %ymm3
; NoVLX-NEXT:    vpternlogd $255, %zmm4, %zmm4, %zmm4 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm4, %xmm4
; NoVLX-NEXT:    vpcmpgtw %ymm3, %ymm1, %ymm1
; NoVLX-NEXT:    vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm3, %xmm3
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vinserti128 $1, %xmm5, %ymm2, %ymm2
; NoVLX-NEXT:    vpcmpgtw %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpand %xmm4, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %1 = bitcast <8 x i64> %__b to <32 x i16>
  %2 = icmp sgt <32 x i16> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtw_v32i1_v64i1_mask_mem(i32 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtw_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtw (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtw_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
; NoVLX-NEXT:    vmovq %xmm1, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm2
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm4
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm3, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm4, %zmm4, %zmm4 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm4, %xmm4
; NoVLX-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
; NoVLX-NEXT:    vpcmpgtw (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpand %xmm4, %xmm0, %xmm0
; NoVLX-NEXT:    vpcmpgtw 32(%rsi), %ymm1, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <32 x i16>
  %2 = icmp sgt <32 x i16> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i8 @test_vpcmpsgtd_v4i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpsgtd_v4i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsgtd_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpsgtd_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpsgtd_v4i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsgtd_v4i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtd_v4i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpsgtd_v4i1_v16i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsgtd_v4i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgtd_v4i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtd_v4i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsgtd_v4i1_v32i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgtd_v4i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtd_v4i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtd_v4i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgtd_v4i1_v64i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpsgtd_v8i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsgtd_v8i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtd_v8i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpsgtd_v8i1_v16i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsgtd_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgtd_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtd_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsgtd_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgtd_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtd_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtd_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgtd_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpsgtd_v16i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgtd_v16i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtd (%rdi), %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtd_v16i1_v32i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd (%rsi), %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsgtd_v16i1_v32i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi){1to16}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtd (%rdi){1to16}, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd (%rsi){1to16}, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %3, %2
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgtd_v16i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtd_v16i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtd (%rdi), %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtd_v16i1_v64i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd (%rsi), %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgtd_v16i1_v64i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtd (%rdi){1to16}, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtd (%rdi){1to16}, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtd (%rsi){1to16}, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %3, %2
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i4 @test_vpcmpsgtq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_vpcmpsgtq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_masked_vpcmpsgtq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}

define zeroext i4 @test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}


define zeroext i4 @test_vpcmpsgtq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}


define zeroext i8 @test_vpcmpsgtq_v2i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpsgtq_v2i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsgtq_v2i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpsgtq_v2i1_v8i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpsgtq_v2i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsgtq_v2i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtq_v2i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpsgtq_v2i1_v16i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsgtq_v2i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgtq_v2i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtq_v2i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsgtq_v2i1_v32i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgtq_v2i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtq_v2i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtq_v2i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgtq_v2i1_v64i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sgt <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i8 @test_vpcmpsgtq_v4i1_v8i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpsgtq_v4i1_v8i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsgtq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpsgtq_v4i1_v8i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpsgtq_v4i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsgtq_v4i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpsgtq_v4i1_v16i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsgtq_v4i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgtq_v4i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsgtq_v4i1_v32i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgtq_v4i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtq_v4i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgtq_v4i1_v64i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpsgtq_v8i1_v16i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsgtq_v8i1_v16i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtq_v8i1_v16i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpsgtq_v8i1_v16i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsgtq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgtq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsgtq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgtq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgtq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgtq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpgtq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sgt <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpsgeb_v16i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeb_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleb %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeb_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp sge <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgeb_v16i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeb_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltb (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeb_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp sge <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgeb_v16i1_v32i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeb_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleb %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeb_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp sge <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgeb_v16i1_v32i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeb_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltb (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeb_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp sge <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgeb_v16i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeb_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleb %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeb_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp sge <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgeb_v16i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeb_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltb (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeb_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp sge <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgeb_v16i1_v64i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeb_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleb %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeb_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp sge <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgeb_v16i1_v64i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeb_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltb (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeb_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp sge <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgeb_v32i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeb_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleb %ymm0, %ymm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeb_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtb %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %1 = bitcast <4 x i64> %__b to <32 x i8>
  %2 = icmp sge <32 x i8> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgeb_v32i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeb_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltb (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeb_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtb %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <32 x i8>
  %2 = icmp sge <32 x i8> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgeb_v32i1_v64i1_mask(i32 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeb_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleb %ymm0, %ymm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeb_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm3, %xmm3
; NoVLX-NEXT:    vpcmpgtb %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %1 = bitcast <4 x i64> %__b to <32 x i8>
  %2 = icmp sge <32 x i8> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgeb_v32i1_v64i1_mask_mem(i32 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeb_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltb (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeb_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm3
; NoVLX-NEXT:    vpcmpgtb %ymm0, %ymm3, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm3, %ymm3, %ymm3
; NoVLX-NEXT:    vpxor %ymm3, %ymm0, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vpand %xmm2, %xmm3, %xmm2
; NoVLX-NEXT:    vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT:    vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT:    vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <32 x i8>
  %2 = icmp sge <32 x i8> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpsgew_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmplew %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsgew_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgew_v8i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmplew %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsgew_v8i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsgew_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmplew %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgew_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgew_v8i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmplew %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgew_v8i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgew_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmplew %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgew_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgew_v8i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmplew %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgew_v8i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp sge <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpsgew_v16i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmplew %ymm0, %ymm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp sge <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgew_v16i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltw (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp sge <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgew_v16i1_v32i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmplew %ymm0, %ymm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp sge <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgew_v16i1_v32i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltw (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp sge <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgew_v16i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmplew %ymm0, %ymm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp sge <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgew_v16i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltw (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp sge <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgew_v16i1_v64i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmplew %ymm0, %ymm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp sge <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgew_v16i1_v64i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltw (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp sge <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgew_v32i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmplew %zmm0, %zmm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vmovq %xmm3, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm2
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm5
; NoVLX-NEXT:    vextracti32x4 $2, %zmm1, %xmm8
; NoVLX-NEXT:    vextracti32x4 $3, %zmm1, %xmm4
; NoVLX-NEXT:    vextracti128 $1, %ymm1, %xmm6
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm7
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm2
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm2, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm7, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm5
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm7, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm6, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm7
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm6, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm6
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm8, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm8, %rax
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
; NoVLX-NEXT:    vinserti128 $1, %xmm5, %ymm7, %ymm3
; NoVLX-NEXT:    vinserti128 $1, %xmm6, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm3, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %1 = bitcast <8 x i64> %__b to <32 x i16>
  %2 = icmp sge <32 x i16> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgew_v32i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgew_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltw (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgew_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm2
; NoVLX-NEXT:    vmovq %xmm2, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm1
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm1, %xmm3
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm1
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm4
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm1, %ymm1
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm2
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm2, %ymm0
; NoVLX-NEXT:    vmovdqa 32(%rdi), %ymm2
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm2, %ymm1
; NoVLX-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <32 x i16>
  %2 = icmp sge <32 x i16> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgew_v32i1_v64i1_mask(i32 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmplew %zmm0, %zmm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm2
; NoVLX-NEXT:    vmovq %xmm2, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm3
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm4
; NoVLX-NEXT:    vextracti128 $1, %ymm1, %xmm8
; NoVLX-NEXT:    vextracti32x4 $2, %zmm1, %xmm5
; NoVLX-NEXT:    vextracti32x4 $3, %zmm1, %xmm7
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm6
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm3, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm9
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm6, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm4
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm6, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm6
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm7, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm7, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm5, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm7
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm5, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm8, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm8, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    vinserti128 $1, %xmm9, %ymm4, %ymm1
; NoVLX-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm0
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vinserti128 $1, %xmm7, %ymm3, %ymm3
; NoVLX-NEXT:    vpternlogd $255, %zmm4, %zmm4, %zmm4 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm4, %xmm4
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm3, %ymm1
; NoVLX-NEXT:    vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm3, %xmm3
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vinserti128 $1, %xmm5, %ymm2, %ymm2
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm2, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpand %xmm4, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %1 = bitcast <8 x i64> %__b to <32 x i16>
  %2 = icmp sge <32 x i16> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgew_v32i1_v64i1_mask_mem(i32 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgew_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltw (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgew_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
; NoVLX-NEXT:    vmovq %xmm1, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm2
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm4
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm3, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm4, %zmm4, %zmm4 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm4, %xmm4
; NoVLX-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm3
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm3, %ymm0
; NoVLX-NEXT:    vmovdqa 32(%rsi), %ymm3
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm3, %ymm1
; NoVLX-NEXT:    vpcmpeqd %ymm3, %ymm3, %ymm3
; NoVLX-NEXT:    vpxor %ymm3, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpand %xmm4, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %ymm3, %ymm1, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <32 x i16>
  %2 = icmp sge <32 x i16> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i8 @test_vpcmpsged_v4i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpled %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpsged_v4i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsged_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpled %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpsged_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpsged_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsged_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpsged_v4i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpled %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsged_v4i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsged_v4i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpled %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsged_v4i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpsged_v4i1_v16i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsged_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsged_v4i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpled %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsged_v4i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsged_v4i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpled %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsged_v4i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsged_v4i1_v32i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsged_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsged_v4i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpled %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsged_v4i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsged_v4i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpled %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsged_v4i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsged_v4i1_v64i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsged_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpsged_v8i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpled %ymm0, %ymm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsged_v8i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsged_v8i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpled %ymm0, %ymm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsged_v8i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpsged_v8i1_v16i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsged_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsged_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpled %ymm0, %ymm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsged_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsged_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpled %ymm0, %ymm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsged_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsged_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsged_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsged_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpled %ymm0, %ymm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsged_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsged_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpled %ymm0, %ymm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsged_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsged_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsged_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpsged_v16i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsged_v16i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpnltd (%rdi), %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsged_v16i1_v32i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsged_v16i1_v32i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpnltd (%rsi), %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsged_v16i1_v32i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi){1to16}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpnltd (%rdi){1to16}, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsged_v16i1_v32i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpnltd (%rsi){1to16}, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %3, %2
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsged_v16i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsged_v16i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpnltd (%rdi), %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsged_v16i1_v64i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsged_v16i1_v64i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpnltd (%rsi), %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsged_v16i1_v64i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltd (%rdi){1to16}, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpnltd (%rdi){1to16}, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsged_v16i1_v64i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpnltd (%rsi){1to16}, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %3, %2
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i4 @test_vpcmpsgeq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_vpcmpsgeq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_masked_vpcmpsgeq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}

define zeroext i4 @test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}


define zeroext i4 @test_vpcmpsgeq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}


define zeroext i8 @test_vpcmpsgeq_v2i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpsgeq_v2i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsgeq_v2i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpsgeq_v2i1_v8i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpsgeq_v2i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsgeq_v2i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgeq_v2i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpsgeq_v2i1_v16i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsgeq_v2i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgeq_v2i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgeq_v2i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsgeq_v2i1_v32i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgeq_v2i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgeq_v2i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgeq_v2i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %xmm0, %xmm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgeq_v2i1_v64i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp sge <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i8 @test_vpcmpsgeq_v4i1_v8i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %ymm0, %ymm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpsgeq_v4i1_v8i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsgeq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %ymm0, %ymm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpsgeq_v4i1_v8i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpsgeq_v4i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %ymm0, %ymm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsgeq_v4i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgeq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %ymm0, %ymm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpsgeq_v4i1_v16i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsgeq_v4i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %ymm0, %ymm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgeq_v4i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgeq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %ymm0, %ymm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsgeq_v4i1_v32i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgeq_v4i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %ymm0, %ymm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgeq_v4i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgeq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %ymm0, %ymm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgeq_v4i1_v64i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpsgeq_v8i1_v16i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpsgeq_v8i1_v16i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpnltq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgeq_v8i1_v16i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpsgeq_v8i1_v16i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpnltq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpsgeq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpsgeq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpnltq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgeq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpsgeq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpnltq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpsgeq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpsgeq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpnltq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgeq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpsgeq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpnltq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpnltq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp sge <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpultb_v16i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultb_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltub %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultb_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp ult <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpultb_v16i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultb_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltub (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultb_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp ult <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultb_v16i1_v32i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultb_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltub %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultb_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp ult <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpultb_v16i1_v32i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultb_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltub (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultb_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp ult <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpultb_v16i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultb_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltub %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultb_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp ult <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultb_v16i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultb_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltub (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultb_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp ult <16 x i8> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultb_v16i1_v64i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultb_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltub %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultb_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %1 = bitcast <2 x i64> %__b to <16 x i8>
  %2 = icmp ult <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultb_v16i1_v64i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultb_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltub (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultb_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <16 x i8>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <16 x i8>
  %2 = icmp ult <16 x i8> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpultb_v32i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultb_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltub %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultb_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtb %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %1 = bitcast <4 x i64> %__b to <32 x i8>
  %2 = icmp ult <32 x i8> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultb_v32i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultb_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltub (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultb_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rdi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtb %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <32 x i8>
  %2 = icmp ult <32 x i8> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultb_v32i1_v64i1_mask(i32 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultb_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltub %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultb_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm3, %xmm3
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %ymm4, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm4, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtb %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %1 = bitcast <4 x i64> %__b to <32 x i8>
  %2 = icmp ult <32 x i8> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultb_v32i1_v64i1_mask_mem(i32 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultb_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltub (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultb_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT:    vpxor %ymm3, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rsi), %ymm3, %ymm3
; NoVLX-NEXT:    vpcmpgtb %ymm0, %ymm3, %ymm0
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vpand %xmm2, %xmm3, %xmm2
; NoVLX-NEXT:    vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT:    vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT:    vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <32 x i8>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <32 x i8>
  %2 = icmp ult <32 x i8> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpultw_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpultw_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultw_v8i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpultw_v8i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpultw_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpultw_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultw_v8i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpultw_v8i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpultw_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultw_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultw_v8i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %1 = bitcast <2 x i64> %__b to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultw_v8i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <8 x i16>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <8 x i16>
  %2 = icmp ult <8 x i16> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpultw_v16i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp ult <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpultw_v16i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rdi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp ult <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultw_v16i1_v32i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp ult <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpultw_v16i1_v32i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rsi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp ult <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpultw_v16i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp ult <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultw_v16i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rdi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp ult <16 x i16> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultw_v16i1_v64i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %1 = bitcast <4 x i64> %__b to <16 x i16>
  %2 = icmp ult <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultw_v16i1_v64i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rsi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <16 x i16>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <16 x i16>
  %2 = icmp ult <16 x i16> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, 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, i32 31, 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, i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpultw_v32i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vmovq %xmm3, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm2
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm5
; NoVLX-NEXT:    vextracti32x4 $2, %zmm1, %xmm8
; NoVLX-NEXT:    vextracti32x4 $3, %zmm1, %xmm4
; NoVLX-NEXT:    vextracti128 $1, %ymm1, %xmm6
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm7
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm2
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm2, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm7, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm5
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm7, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm6, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm7
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm6, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm6
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm8, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm8, %rax
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
; NoVLX-NEXT:    vinserti128 $1, %xmm5, %ymm7, %ymm3
; NoVLX-NEXT:    vinserti128 $1, %xmm6, %ymm1, %ymm1
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm4
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm4, %xmm4
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vinserti128 $1, %xmm2, %ymm4, %ymm2
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm4 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm4, %ymm3, %ymm3
; NoVLX-NEXT:    vpxor %ymm4, %ymm2, %ymm2
; NoVLX-NEXT:    vpcmpgtw %ymm3, %ymm2, %ymm2
; NoVLX-NEXT:    vpmovsxwd %ymm2, %zmm2
; NoVLX-NEXT:    vpmovdb %zmm2, %xmm2
; NoVLX-NEXT:    vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT:    vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %ymm4, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm4, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %1 = bitcast <8 x i64> %__b to <32 x i16>
  %2 = icmp ult <32 x i16> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultw_v32i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultw_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuw (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultw_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm2
; NoVLX-NEXT:    vmovq %xmm2, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm1
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm1, %xmm3
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm1
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm4
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm1, %ymm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor 32(%rdi), %ymm2, %ymm3
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm3, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rdi), %ymm2, %ymm1
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <32 x i16>
  %2 = icmp ult <32 x i16> %0, %1
  %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultw_v32i1_v64i1_mask(i32 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v32i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v32i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm2
; NoVLX-NEXT:    vmovq %xmm2, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm3
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm4
; NoVLX-NEXT:    vextracti128 $1, %ymm1, %xmm8
; NoVLX-NEXT:    vextracti32x4 $2, %zmm1, %xmm5
; NoVLX-NEXT:    vextracti32x4 $3, %zmm1, %xmm7
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm6
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm2, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm3, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm9
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm6, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm4
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm6, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm6
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vmovq %xmm7, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm7, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm5, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm7
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm3
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpextrq $1, %xmm5, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm8, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm5
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpextrq $1, %xmm8, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm5, %xmm5
; NoVLX-NEXT:    vmovq %xmm1, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm5, %xmm5
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    vinserti128 $1, %xmm9, %ymm4, %ymm1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm0
; NoVLX-NEXT:    vpternlogd $255, %zmm4, %zmm4, %zmm4 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm4, %xmm4
; NoVLX-NEXT:    vinserti128 $1, %xmm7, %ymm3, %ymm3
; NoVLX-NEXT:    vpternlogd $255, %zmm6, %zmm6, %zmm6 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm6, %xmm6
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vinserti128 $1, %xmm5, %ymm2, %ymm2
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm5 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm5, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm5, %ymm2, %ymm2
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm2, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpand %xmm4, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %ymm5, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor %ymm5, %ymm3, %ymm2
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm2, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm6, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %1 = bitcast <8 x i64> %__b to <32 x i16>
  %2 = icmp ult <32 x i16> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultw_v32i1_v64i1_mask_mem(i32 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultw_v32i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuw (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultw_v32i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $96, %rsp
; NoVLX-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
; NoVLX-NEXT:    vmovq %xmm1, %rax
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    movq %rax, %rdx
; NoVLX-NEXT:    vmovd %eax, %xmm2
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT:    vextracti32x4 $2, %zmm0, %xmm4
; NoVLX-NEXT:    shrq $32, %rdx
; NoVLX-NEXT:    vpinsrw $2, %edx, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm1, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm1
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vmovq %xmm4, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm2
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpextrq $1, %xmm4, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT:    vmovq %xmm3, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm3, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT:    vmovq %xmm0, %rcx
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT:    movl %ecx, %eax
; NoVLX-NEXT:    shrl $16, %eax
; NoVLX-NEXT:    vmovd %ecx, %xmm4
; NoVLX-NEXT:    vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    movq %rcx, %rax
; NoVLX-NEXT:    shrq $32, %rax
; NoVLX-NEXT:    vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT:    vpextrq $1, %xmm0, %rax
; NoVLX-NEXT:    shrq $48, %rcx
; NoVLX-NEXT:    vpinsrw $3, %ecx, %xmm4, %xmm0
; NoVLX-NEXT:    movl %eax, %ecx
; NoVLX-NEXT:    shrl $16, %ecx
; NoVLX-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    movq %rax, %rcx
; NoVLX-NEXT:    shrq $32, %rcx
; NoVLX-NEXT:    vpinsrw $6, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT:    vpternlogd $255, %zmm4, %zmm4, %zmm4 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm4, %xmm4
; NoVLX-NEXT:    vpternlogd $255, %zmm5, %zmm5, %zmm5 {%k2} {z}
; NoVLX-NEXT:    vpmovdb %zmm5, %xmm5
; NoVLX-NEXT:    shrq $48, %rax
; NoVLX-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
; NoVLX-NEXT:    vpinsrw $7, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
; NoVLX-NEXT:    vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rsi), %ymm2, %ymm3
; NoVLX-NEXT:    vpcmpgtw %ymm0, %ymm3, %ymm0
; NoVLX-NEXT:    vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpand %xmm4, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpxor 32(%rsi), %ymm2, %ymm2
; NoVLX-NEXT:    vpcmpgtw %ymm1, %ymm2, %ymm1
; NoVLX-NEXT:    vpmovsxwd %ymm1, %zmm1
; NoVLX-NEXT:    vpmovdb %zmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm5, %xmm1, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %ecx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    shlq $32, %rax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <32 x i16>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <32 x i16>
  %2 = icmp ult <32 x i16> %0, %1
  %3 = bitcast i32 %__u to <32 x i1>
  %4 = and <32 x i1> %2, %3
  %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i8 @test_vpcmpultd_v4i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpultd_v4i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpultd_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpultd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpultd_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpultd_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpultd_v4i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpultd_v4i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultd_v4i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpultd_v4i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpultd_v4i1_v16i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultd_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpultd_v4i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpultd_v4i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultd_v4i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpultd_v4i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpultd_v4i1_v32i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultd_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpultd_v4i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultd_v4i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultd_v4i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %1 = bitcast <2 x i64> %__b to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultd_v4i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x i32>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpultd_v4i1_v64i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultd_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastd (%rsi), %xmm1
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <4 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpultd_v8i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpultd_v8i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultd_v8i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpultd_v8i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpultd_v8i1_v16i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultd_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpultd_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpultd_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultd_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpultd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpultd_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpultd_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultd_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultd_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %1 = bitcast <4 x i64> %__b to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x i32>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpultd_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <8 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i32> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i32 @test_vpcmpultd_v16i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpultd_v16i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpltud (%rdi), %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultd_v16i1_v32i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpultd_v16i1_v32i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud (%rsi), %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpultd_v16i1_v32i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi){1to16}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpltud (%rdi){1to16}, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultd_v16i1_v32i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud (%rsi){1to16}, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %3, %2
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpultd_v16i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultd_v16i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpltud (%rdi), %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultd_v16i1_v64i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %1 = bitcast <8 x i64> %__b to <16 x i32>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultd_v16i1_v64i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud (%rsi), %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x i32>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpultd_v16i1_v64i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltud (%rdi){1to16}, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpltud (%rdi){1to16}, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultd_v16i1_v64i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltud (%rsi){1to16}, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x i32>
  %load = load i32, i32* %__b
  %vec = insertelement <16 x i32> undef, i32 %load, i32 0
  %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <16 x i32> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %3, %2
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i4 @test_vpcmpultq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_vpcmpultq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_masked_vpcmpultq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}

define zeroext i4 @test_masked_vpcmpultq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}


define zeroext i4 @test_vpcmpultq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_masked_vpcmpultq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}


define zeroext i8 @test_vpcmpultq_v2i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpultq_v2i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpultq_v2i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpultq_v2i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpultq_v2i1_v8i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpultq_v2i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpultq_v2i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpultq_v2i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultq_v2i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpultq_v2i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpultq_v2i1_v16i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultq_v2i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpultq_v2i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpultq_v2i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultq_v2i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpultq_v2i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpultq_v2i1_v32i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultq_v2i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpultq_v2i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultq_v2i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultq_v2i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %1 = bitcast <2 x i64> %__b to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultq_v2i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x i64>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %2, %extract.i
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpultq_v2i1_v64i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultq_v2i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %xmm1
; NoVLX-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kmovw %k0, %ecx
; NoVLX-NEXT:    vmovd %ecx, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <2 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %2 = icmp ult <2 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
  %4 = and <2 x i1> %extract.i, %2
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i8 @test_vpcmpultq_v4i1_v8i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vpcmpultq_v4i1_v8i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rdi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpultq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vpcmpultq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rsi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i8 @test_vpcmpultq_v4i1_v8i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vpcmpultq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}


define zeroext i16 @test_vpcmpultq_v4i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpultq_v4i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rdi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpultq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rsi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpultq_v4i1_v16i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpultq_v4i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpultq_v4i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rdi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpultq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rsi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpultq_v4i1_v32i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpultq_v4i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultq_v4i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rdi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %1 = bitcast <4 x i64> %__b to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor (%rsi), %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x i64>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %2, %extract.i
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpultq_v4i1_v64i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpbroadcastq (%rsi), %ymm1
; NoVLX-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT:    vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kmovw %edi, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edx
; NoVLX-NEXT:    kmovw %k0, %esi
; NoVLX-NEXT:    vmovd %esi, %xmm1
; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <4 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <4 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = and <4 x i1> %extract.i, %2
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i16 @test_vpcmpultq_v8i1_v16i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vpcmpultq_v8i1_v16i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpltuq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultq_v8i1_v16i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vpcmpultq_v8i1_v16i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i16 @test_vpcmpultq_v8i1_v16i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vpcmpltuq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vpcmpultq_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}


define zeroext i32 @test_vpcmpultq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vpcmpultq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpltuq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vpcmpultq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i32 @test_vpcmpultq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vpcmpltuq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vpcmpultq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}


define zeroext i64 @test_vpcmpultq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vpcmpultq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpltuq (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %1 = bitcast <8 x i64> %__b to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vpcmpultq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x i64>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


define zeroext i64 @test_vpcmpultq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vpcmpltuq (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vpcmpltuq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vpcmpultq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x i64>
  %load = load i64, i64* %__b
  %vec = insertelement <8 x i64> undef, i64 %load, i32 0
  %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = icmp ult <8 x i64> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %3, %2
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}


declare i16 @llvm.x86.avx512.mask.cmp.ps.512(<16 x float> , <16 x float> , i32, i16, i32)
define zeroext i8 @test_vcmpoeqps_v4i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %1 = bitcast <2 x i64> %__b to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vcmpoeqps_v4i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqps (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vcmpoeqps_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vbroadcastss (%rdi), %xmm1
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load float, float* %__b
  %vec = insertelement <4 x float> undef, float %load, i32 0
  %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vcmpoeqps_v4i1_v8i1_mask(i4 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %1 = bitcast <2 x i64> %__b to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vcmpoeqps_v4i1_v8i1_mask_mem(i4 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqps (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vcmpoeqps_v4i1_v8i1_mask_mem_b(i4 zeroext %__u, <2 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vbroadcastss (%rsi), %xmm2
; NoVLX-NEXT:    vcmpeqps %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load float, float* %__b
  %vec = insertelement <4 x float> undef, float %load, i32 0
  %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}



define zeroext i16 @test_vcmpoeqps_v4i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %1 = bitcast <2 x i64> %__b to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vcmpoeqps_v4i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqps (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vcmpoeqps_v4i1_v16i1_mask_mem_b(<2 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vbroadcastss (%rdi), %xmm1
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load float, float* %__b
  %vec = insertelement <4 x float> undef, float %load, i32 0
  %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vcmpoeqps_v4i1_v16i1_mask(i4 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %1 = bitcast <2 x i64> %__b to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vcmpoeqps_v4i1_v16i1_mask_mem(i4 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqps (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vcmpoeqps_v4i1_v16i1_mask_mem_b(i4 zeroext %__u, <2 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vbroadcastss (%rsi), %xmm2
; NoVLX-NEXT:    vcmpeqps %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load float, float* %__b
  %vec = insertelement <4 x float> undef, float %load, i32 0
  %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}



define zeroext i32 @test_vcmpoeqps_v4i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %1 = bitcast <2 x i64> %__b to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqps_v4i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqps (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqps_v4i1_v32i1_mask_mem_b(<2 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vbroadcastss (%rdi), %xmm1
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load float, float* %__b
  %vec = insertelement <4 x float> undef, float %load, i32 0
  %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vcmpoeqps_v4i1_v32i1_mask(i4 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %1 = bitcast <2 x i64> %__b to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqps_v4i1_v32i1_mask_mem(i4 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqps (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqps_v4i1_v32i1_mask_mem_b(i4 zeroext %__u, <2 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vbroadcastss (%rsi), %xmm2
; NoVLX-NEXT:    vcmpeqps %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load float, float* %__b
  %vec = insertelement <4 x float> undef, float %load, i32 0
  %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}



define zeroext i64 @test_vcmpoeqps_v4i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %1 = bitcast <2 x i64> %__b to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqps_v4i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqps (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqps_v4i1_v64i1_mask_mem_b(<2 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi){1to4}, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vbroadcastss (%rdi), %xmm1
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load float, float* %__b
  %vec = insertelement <4 x float> undef, float %load, i32 0
  %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vcmpoeqps_v4i1_v64i1_mask(i4 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %1 = bitcast <2 x i64> %__b to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqps_v4i1_v64i1_mask_mem(i4 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqps (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <4 x float>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqps_v4i1_v64i1_mask_mem_b(i4 zeroext %__u, <2 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi){1to4}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vbroadcastss (%rsi), %xmm2
; NoVLX-NEXT:    vcmpeqps %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vandps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <4 x float>
  %load = load float, float* %__b
  %vec = insertelement <4 x float> undef, float %load, i32 0
  %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x float> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}



define zeroext i16 @test_vcmpoeqps_v8i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %1 = bitcast <4 x i64> %__b to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vcmpoeqps_v8i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovaps (%rdi), %ymm1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vcmpoeqps_v8i1_v16i1_mask_mem_b(<4 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vbroadcastss (%rdi), %ymm1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load float, float* %__b
  %vec = insertelement <8 x float> undef, float %load, i32 0
  %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vcmpoeqps_v8i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %1 = bitcast <4 x i64> %__b to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vcmpoeqps_v8i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovaps (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vcmpoeqps_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vbroadcastss (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftlw $8, %k0, %k0
; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load float, float* %__b
  %vec = insertelement <8 x float> undef, float %load, i32 0
  %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}



define zeroext i32 @test_vcmpoeqps_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %1 = bitcast <4 x i64> %__b to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqps_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovaps (%rdi), %ymm1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqps_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vbroadcastss (%rdi), %ymm1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load float, float* %__b
  %vec = insertelement <8 x float> undef, float %load, i32 0
  %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vcmpoeqps_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %1 = bitcast <4 x i64> %__b to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqps_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovaps (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqps_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vbroadcastss (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load float, float* %__b
  %vec = insertelement <8 x float> undef, float %load, i32 0
  %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}



define zeroext i64 @test_vcmpoeqps_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %1 = bitcast <4 x i64> %__b to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqps_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovaps (%rdi), %ymm1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqps_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi){1to8}, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vbroadcastss (%rdi), %ymm1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load float, float* %__b
  %vec = insertelement <8 x float> undef, float %load, i32 0
  %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vcmpoeqps_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %1 = bitcast <4 x i64> %__b to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqps_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vmovaps (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <8 x float>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqps_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi){1to8}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
; NoVLX-NEXT:    vbroadcastss (%rsi), %ymm1
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <8 x float>
  %load = load float, float* %__b
  %vec = insertelement <8 x float> undef, float %load, i32 0
  %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x float> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}



define zeroext i32 @test_vcmpoeqps_v16i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %1 = bitcast <8 x i64> %__b to <16 x float>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqps_v16i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqps (%rdi), %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x float>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqps_v16i1_v32i1_mask_mem_b(<8 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi){1to16}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqps (%rdi){1to16}, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %load = load float, float* %__b
  %vec = insertelement <16 x float> undef, float %load, i32 0
  %1 = shufflevector <16 x float> %vec, <16 x float> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vcmpoeqps_v16i1_v32i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %1 = bitcast <8 x i64> %__b to <16 x float>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqps_v16i1_v32i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps (%rsi), %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x float>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqps_v16i1_v32i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps (%rsi){1to16}, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %load = load float, float* %__b
  %vec = insertelement <16 x float> undef, float %load, i32 0
  %1 = shufflevector <16 x float> %vec, <16 x float> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}



define zeroext i32 @test_vcmpoeqps_v16i1_v32i1_sae_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; CHECK-LABEL: test_vcmpoeqps_v16i1_v32i1_sae_mask:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    vcmpleps {sae}, %zmm1, %zmm0, %k0
; CHECK-NEXT:    kmovw %k0, %eax
; CHECK-NEXT:    vzeroupper
; CHECK-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %1 = bitcast <8 x i64> %__b to <16 x float>
  %2 = call i16 @llvm.x86.avx512.mask.cmp.ps.512(<16 x float> %0, <16 x float> %1, i32 2, i16 -1, i32 8)
  %3 = zext i16 %2 to i32
  ret i32 %3
}

define zeroext i32 @test_masked_vcmpoeqps_v16i1_v32i1_sae_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_sae_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpleps {sae}, %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovw %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_sae_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpleps {sae}, %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %1 = bitcast <8 x i64> %__b to <16 x float>
  %2 = call i16 @llvm.x86.avx512.mask.cmp.ps.512(<16 x float> %0, <16 x float> %1, i32 2, i16 %__u, i32 8)
  %3 = zext i16 %2 to i32
  ret i32 %3
}



define zeroext i64 @test_vcmpoeqps_v16i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %1 = bitcast <8 x i64> %__b to <16 x float>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqps_v16i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqps (%rdi), %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x float>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqps_v16i1_v64i1_mask_mem_b(<8 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqps (%rdi){1to16}, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqps (%rdi){1to16}, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %load = load float, float* %__b
  %vec = insertelement <16 x float> undef, float %load, i32 0
  %1 = shufflevector <16 x float> %vec, <16 x float> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vcmpoeqps_v16i1_v64i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %1 = bitcast <8 x i64> %__b to <16 x float>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqps_v16i1_v64i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps (%rsi), %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <16 x float>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqps_v16i1_v64i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, float* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqps (%rsi){1to16}, %zmm0, %k1 {%k1}
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %load = load float, float* %__b
  %vec = insertelement <16 x float> undef, float %load, i32 0
  %1 = shufflevector <16 x float> %vec, <16 x float> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <16 x float> %0, %1
  %3 = bitcast i16 %__u to <16 x i1>
  %4 = and <16 x i1> %2, %3
  %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,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,i32 31,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,i32 31,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,i32 31>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}



define zeroext i64 @test_vcmpoeqps_v16i1_v64i1_sae_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqps_v16i1_v64i1_sae_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpleps {sae}, %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    movzwl %ax, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqps_v16i1_v64i1_sae_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpleps {sae}, %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movzwl %ax, %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %1 = bitcast <8 x i64> %__b to <16 x float>
  %2 = call i16 @llvm.x86.avx512.mask.cmp.ps.512(<16 x float> %0, <16 x float> %1, i32 2, i16 -1, i32 8)
  %3 = zext i16 %2 to i64
  ret i64 %3
}

define zeroext i64 @test_masked_vcmpoeqps_v16i1_v64i1_sae_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_sae_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpleps {sae}, %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    movzwl %ax, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_sae_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpleps {sae}, %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movzwl %ax, %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <16 x float>
  %1 = bitcast <8 x i64> %__b to <16 x float>
  %2 = call i16 @llvm.x86.avx512.mask.cmp.ps.512(<16 x float> %0, <16 x float> %1, i32 2, i16 %__u, i32 8)
  %3 = zext i16 %2 to i64
  ret i64 %3
}



declare i8 @llvm.x86.avx512.mask.cmp.pd.512(<8 x double> , <8 x double> , i32, i8, i32)
define zeroext i4 @test_vcmpoeqpd_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %1 = bitcast <2 x i64> %__b to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_vcmpoeqpd_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_vcmpoeqpd_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovddup {{.*#+}} xmm1 = mem[0,0]
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load double, double* %__b
  %vec = insertelement <2 x double> undef, double %load, i32 0
  %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %4 = bitcast <4 x i1> %3 to i4
  ret i4 %4
}

define zeroext i4 @test_masked_vcmpoeqpd_v2i1_v4i1_mask(i2 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %1 = bitcast <2 x i64> %__b to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}

define zeroext i4 @test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem(i2 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}

define zeroext i4 @test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem_b(i2 zeroext %__u, <2 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vmovddup {{.*#+}} xmm2 = mem[0,0]
; NoVLX-NEXT:    vcmpeqpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,2],zero,zero
; NoVLX-NEXT:    vpslld $31, %ymm0, %ymm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movb %al, -{{[0-9]+}}(%rsp)
; NoVLX-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load double, double* %__b
  %vec = insertelement <2 x double> undef, double %load, i32 0
  %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %6 = bitcast <4 x i1> %5 to i4
  ret i4 %6
}



define zeroext i8 @test_vcmpoeqpd_v2i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %1 = bitcast <2 x i64> %__b to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vcmpoeqpd_v2i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vcmpoeqpd_v2i1_v8i1_mask_mem_b(<2 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovddup {{.*#+}} xmm1 = mem[0,0]
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load double, double* %__b
  %vec = insertelement <2 x double> undef, double %load, i32 0
  %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vcmpoeqpd_v2i1_v8i1_mask(i2 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %1 = bitcast <2 x i64> %__b to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem(i2 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem_b(i2 zeroext %__u, <2 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vmovddup {{.*#+}} xmm2 = mem[0,0]
; NoVLX-NEXT:    vcmpeqpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load double, double* %__b
  %vec = insertelement <2 x double> undef, double %load, i32 0
  %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}



define zeroext i16 @test_vcmpoeqpd_v2i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %1 = bitcast <2 x i64> %__b to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vcmpoeqpd_v2i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vcmpoeqpd_v2i1_v16i1_mask_mem_b(<2 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vmovddup {{.*#+}} xmm1 = mem[0,0]
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load double, double* %__b
  %vec = insertelement <2 x double> undef, double %load, i32 0
  %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vcmpoeqpd_v2i1_v16i1_mask(i2 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %1 = bitcast <2 x i64> %__b to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem(i2 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem_b(i2 zeroext %__u, <2 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vmovddup {{.*#+}} xmm2 = mem[0,0]
; NoVLX-NEXT:    vcmpeqpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load double, double* %__b
  %vec = insertelement <2 x double> undef, double %load, i32 0
  %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}



define zeroext i32 @test_vcmpoeqpd_v2i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %1 = bitcast <2 x i64> %__b to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqpd_v2i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqpd_v2i1_v32i1_mask_mem_b(<2 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vmovddup {{.*#+}} xmm1 = mem[0,0]
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load double, double* %__b
  %vec = insertelement <2 x double> undef, double %load, i32 0
  %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vcmpoeqpd_v2i1_v32i1_mask(i2 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %1 = bitcast <2 x i64> %__b to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem(i2 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem_b(i2 zeroext %__u, <2 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vmovddup {{.*#+}} xmm2 = mem[0,0]
; NoVLX-NEXT:    vcmpeqpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load double, double* %__b
  %vec = insertelement <2 x double> undef, double %load, i32 0
  %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}



define zeroext i64 @test_vcmpoeqpd_v2i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %1 = bitcast <2 x i64> %__b to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqpd_v2i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqpd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqpd_v2i1_v64i1_mask_mem_b(<2 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to2}, %xmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vmovddup {{.*#+}} xmm1 = mem[0,0]
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load double, double* %__b
  %vec = insertelement <2 x double> undef, double %load, i32 0
  %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vcmpoeqpd_v2i1_v64i1_mask(i2 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %1 = bitcast <2 x i64> %__b to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem(i2 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load <2 x i64>, <2 x i64>* %__b
  %1 = bitcast <2 x i64> %load to <2 x double>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem_b(i2 zeroext %__u, <2 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vmovddup {{.*#+}} xmm2 = mem[0,0]
; NoVLX-NEXT:    vcmpeqpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <2 x i64> %__a to <2 x double>
  %load = load double, double* %__b
  %vec = insertelement <2 x double> undef, double %load, i32 0
  %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
  %2 = fcmp oeq <2 x double> %0, %1
  %3 = bitcast i2 %__u to <2 x i1>
  %4 = and <2 x i1> %2, %3
  %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}



define zeroext i8 @test_vcmpoeqpd_v4i1_v8i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %1 = bitcast <4 x i64> %__b to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vcmpoeqpd_v4i1_v8i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_vcmpoeqpd_v4i1_v8i1_mask_mem_b(<4 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vbroadcastsd (%rdi), %ymm1
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load double, double* %__b
  %vec = insertelement <4 x double> undef, double %load, i32 0
  %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <8 x i1> %3 to i8
  ret i8 %4
}

define zeroext i8 @test_masked_vcmpoeqpd_v4i1_v8i1_mask(i4 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %1 = bitcast <4 x i64> %__b to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem(i4 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}

define zeroext i8 @test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem_b(i4 zeroext %__u, <4 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vbroadcastsd (%rsi), %ymm2
; NoVLX-NEXT:    vcmpeqpd %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load double, double* %__b
  %vec = insertelement <4 x double> undef, double %load, i32 0
  %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <8 x i1> %5 to i8
  ret i8 %6
}



define zeroext i16 @test_vcmpoeqpd_v4i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %1 = bitcast <4 x i64> %__b to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vcmpoeqpd_v4i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vcmpoeqpd_v4i1_v16i1_mask_mem_b(<4 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vbroadcastsd (%rdi), %ymm1
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load double, double* %__b
  %vec = insertelement <4 x double> undef, double %load, i32 0
  %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vcmpoeqpd_v4i1_v16i1_mask(i4 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %1 = bitcast <4 x i64> %__b to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem(i4 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem_b(i4 zeroext %__u, <4 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vbroadcastsd (%rsi), %ymm2
; NoVLX-NEXT:    vcmpeqpd %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    andl $1, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
; NoVLX-NEXT:    korw %k1, %k2, %k1
; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load double, double* %__b
  %vec = insertelement <4 x double> undef, double %load, i32 0
  %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}



define zeroext i32 @test_vcmpoeqpd_v4i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %1 = bitcast <4 x i64> %__b to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqpd_v4i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqpd (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqpd_v4i1_v32i1_mask_mem_b(<4 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vbroadcastsd (%rdi), %ymm1
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load double, double* %__b
  %vec = insertelement <4 x double> undef, double %load, i32 0
  %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vcmpoeqpd_v4i1_v32i1_mask(i4 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %1 = bitcast <4 x i64> %__b to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem(i4 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem_b(i4 zeroext %__u, <4 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vbroadcastsd (%rsi), %ymm2
; NoVLX-NEXT:    vcmpeqpd %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load double, double* %__b
  %vec = insertelement <4 x double> undef, double %load, i32 0
  %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}



define zeroext i64 @test_vcmpoeqpd_v4i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %1 = bitcast <4 x i64> %__b to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqpd_v4i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqpd (%rdi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqpd_v4i1_v64i1_mask_mem_b(<4 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to4}, %ymm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vbroadcastsd (%rdi), %ymm1
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load double, double* %__b
  %vec = insertelement <4 x double> undef, double %load, i32 0
  %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vcmpoeqpd_v4i1_v64i1_mask(i4 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %1 = bitcast <4 x i64> %__b to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem(i4 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vcmpeqpd (%rsi), %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load <4 x i64>, <4 x i64>* %__b
  %1 = bitcast <4 x i64> %load to <4 x double>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem_b(i4 zeroext %__u, <4 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to4}, %ymm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT:    vbroadcastsd (%rsi), %ymm2
; NoVLX-NEXT:    vcmpeqpd %ymm2, %ymm0, %ymm0
; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT:    vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT:    vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT:    vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <4 x i64> %__a to <4 x double>
  %load = load double, double* %__b
  %vec = insertelement <4 x double> undef, double %load, i32 0
  %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <4 x double> %0, %1
  %3 = bitcast i4 %__u to <4 x i1>
  %4 = and <4 x i1> %2, %3
  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}



define zeroext i16 @test_vcmpoeqpd_v8i1_v16i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vcmpoeqpd_v8i1_v16i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_vcmpoeqpd_v8i1_v16i1_mask_mem_b(<8 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmpeqpd (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load double, double* %__b
  %vec = insertelement <8 x double> undef, double %load, i32 0
  %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <16 x i1> %3 to i16
  ret i16 %4
}

define zeroext i16 @test_masked_vcmpoeqpd_v8i1_v16i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}

define zeroext i16 @test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load double, double* %__b
  %vec = insertelement <8 x double> undef, double %load, i32 0
  %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <16 x i1> %5 to i16
  ret i16 %6
}



define zeroext i16 @test_vcmpoeqpd_v8i1_v16i1_sae_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_sae_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    movzbl %al, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_sae_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movzbl %al, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = call i8 @llvm.x86.avx512.mask.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i8 -1, i32 8)
  %3 = zext i8 %2 to i16
  ret i16 %3
}

define zeroext i16 @test_masked_vcmpoeqpd_v8i1_v16i1_sae_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_sae_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    movzbl %al, %eax
; VLX-NEXT:    # kill: def %ax killed %ax killed %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_sae_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movzbl %al, %eax
; NoVLX-NEXT:    # kill: def %ax killed %ax killed %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = call i8 @llvm.x86.avx512.mask.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i8 %__u, i32 8)
  %3 = zext i8 %2 to i16
  ret i16 %3
}



define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqpd (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpeqpd (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load double, double* %__b
  %vec = insertelement <8 x double> undef, double %load, i32 0
  %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <32 x i1> %3 to i32
  ret i32 %4
}

define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}

define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load double, double* %__b
  %vec = insertelement <8 x double> undef, double %load, i32 0
  %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <32 x i1> %5 to i32
  ret i32 %6
}



define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_sae_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_sae_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_sae_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movzbl %al, %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = call i8 @llvm.x86.avx512.mask.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i8 -1, i32 8)
  %3 = zext i8 %2 to i32
  ret i32 %3
}

define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_sae_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_sae_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovb %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_sae_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movzbl %al, %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = call i8 @llvm.x86.avx512.mask.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i8 %__u, i32 8)
  %3 = zext i8 %2 to i32
  ret i32 %3
}



define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi), %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqpd (%rdi), %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmpeqpd (%rdi){1to8}, %zmm0, %k0
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    vcmpeqpd (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load double, double* %__b
  %vec = insertelement <8 x double> undef, double %load, i32 0
  %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %4 = bitcast <64 x i1> %3 to i64
  ret i64 %4
}

define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load <8 x i64>, <8 x i64>* %__b
  %1 = bitcast <8 x i64> %load to <8 x double>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}

define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, double* %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem_b:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovq %k0, %rax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem_b:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $64, %rsp
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r8d
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r9d
; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %r10d
; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %esi
; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %edi
; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %eax
; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
; NoVLX-NEXT:    kmovw %k1, %ecx
; NoVLX-NEXT:    kmovw %k0, %edx
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl {{[0-9]+}}(%rsp), %ecx
; NoVLX-NEXT:    shlq $32, %rcx
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    orq %rcx, %rax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %load = load double, double* %__b
  %vec = insertelement <8 x double> undef, double %load, i32 0
  %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %2 = fcmp oeq <8 x double> %0, %1
  %3 = bitcast i8 %__u to <8 x i1>
  %4 = and <8 x i1> %2, %3
  %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
  %6 = bitcast <64 x i1> %5 to i64
  ret i64 %6
}



define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_sae_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_sae_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    movzbl %al, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_sae_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movzbl %al, %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = call i8 @llvm.x86.avx512.mask.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i8 -1, i32 8)
  %3 = zext i8 %2 to i64
  ret i64 %3
}

define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_sae_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_sae_mask:
; VLX:       # %bb.0: # %entry
; VLX-NEXT:    kmovd %edi, %k1
; VLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0 {%k1}
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    movzbl %al, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_sae_mask:
; NoVLX:       # %bb.0: # %entry
; NoVLX-NEXT:    kmovw %edi, %k1
; NoVLX-NEXT:    vcmplepd {sae}, %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    movzbl %al, %eax
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
entry:
  %0 = bitcast <8 x i64> %__a to <8 x double>
  %1 = bitcast <8 x i64> %__b to <8 x double>
  %2 = call i8 @llvm.x86.avx512.mask.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i8 %__u, i32 8)
  %3 = zext i8 %2 to i64
  ret i64 %3
}

; Test that we understand that cmpps with rounding zeros the upper bits of the mask register.
define i32 @test_cmpm_rnd_zero(<16 x float> %a, <16 x float> %b) {
; VLX-LABEL: test_cmpm_rnd_zero:
; VLX:       # %bb.0:
; VLX-NEXT:    vcmpleps {sae}, %zmm1, %zmm0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    vzeroupper
; VLX-NEXT:    retq
;
; NoVLX-LABEL: test_cmpm_rnd_zero:
; NoVLX:       # %bb.0:
; NoVLX-NEXT:    pushq %rbp
; NoVLX-NEXT:    .cfi_def_cfa_offset 16
; NoVLX-NEXT:    .cfi_offset %rbp, -16
; NoVLX-NEXT:    movq %rsp, %rbp
; NoVLX-NEXT:    .cfi_def_cfa_register %rbp
; NoVLX-NEXT:    andq $-32, %rsp
; NoVLX-NEXT:    subq $32, %rsp
; NoVLX-NEXT:    vcmpleps {sae}, %zmm1, %zmm0, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k0
; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT:    vpmovdb %zmm0, %xmm0
; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT:    kmovw %k0, (%rsp)
; NoVLX-NEXT:    movl (%rsp), %eax
; NoVLX-NEXT:    movq %rbp, %rsp
; NoVLX-NEXT:    popq %rbp
; NoVLX-NEXT:    vzeroupper
; NoVLX-NEXT:    retq
  %res = call i16 @llvm.x86.avx512.mask.cmp.ps.512(<16 x float> %a, <16 x float> %b, i32 2, i16 -1, i32 8)
  %cast = bitcast i16 %res to <16 x i1>
  %shuffle = shufflevector <16 x i1> %cast, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, 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, i32 31>
  %cast2 = bitcast <32 x i1> %shuffle to i32
  ret i32 %cast2
}

define i8 @mask_zero_lower(<4 x i32> %a) {
; VLX-LABEL: mask_zero_lower:
; VLX:       # %bb.0:
; VLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
; VLX-NEXT:    vpcmpltud %xmm1, %xmm0, %k0
; VLX-NEXT:    kshiftlb $4, %k0, %k0
; VLX-NEXT:    kmovd %k0, %eax
; VLX-NEXT:    # kill: def %al killed %al killed %eax
; VLX-NEXT:    retq
;
; NoVLX-LABEL: mask_zero_lower:
; NoVLX:       # %bb.0:
; NoVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k0
; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k1
; NoVLX-NEXT:    kxorw %k0, %k0, %k2
; NoVLX-NEXT:    kshiftrw $4, %k2, %k3
; NoVLX-NEXT:    kxorw %k1, %k3, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $11, %k1, %k1
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftrw $5, %k1, %k2
; NoVLX-NEXT:    kxorw %k0, %k2, %k0
; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
; NoVLX-NEXT:    kshiftrw $10, %k0, %k0
; NoVLX-NEXT:    kxorw %k1, %k0, %k0
; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $9, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
; NoVLX-NEXT:    kmovw %eax, %k2
; NoVLX-NEXT:    kxorw %k2, %k1, %k1
; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
; NoVLX-NEXT:    kshiftrw $8, %k1, %k1
; NoVLX-NEXT:    kxorw %k0, %k1, %k0
; NoVLX-NEXT:    kmovw %k0, %eax
; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
; NoVLX-NEXT:    retq
  %cmp = icmp ult <4 x i32> %a, zeroinitializer
  %concat = shufflevector <4 x i1> %cmp, <4 x i1> zeroinitializer, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3>
  %cast = bitcast <8 x i1> %concat to i8
  ret i8 %cast
}