; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+avx512f,+avx512vl,+avx512bw %s -o - | FileCheck %s

define <16 x i8> @test_i8_to_16(i8 %s) {
; CHECK-LABEL: test_i8_to_16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastb %edi, %xmm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %res = shufflevector <2 x i8> %vec, <2 x i8> 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>
  ret <16 x i8> %res
}
define <16 x i8> @test_masked_i8_to_16_mask0(i8 %s, <16 x i8> %default, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_16_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> %default
  ret <16 x i8> %res
}

define <16 x i8> @test_masked_z_i8_to_16_mask0(i8 %s, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_16_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> zeroinitializer
  ret <16 x i8> %res
}
define <16 x i8> @test_masked_i8_to_16_mask1(i8 %s, <16 x i8> %default, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_16_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> %default
  ret <16 x i8> %res
}

define <16 x i8> @test_masked_z_i8_to_16_mask1(i8 %s, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_16_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> zeroinitializer
  ret <16 x i8> %res
}
define <16 x i8> @test_masked_i8_to_16_mask2(i8 %s, <16 x i8> %default, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_16_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> %default
  ret <16 x i8> %res
}

define <16 x i8> @test_masked_z_i8_to_16_mask2(i8 %s, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_16_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> zeroinitializer
  ret <16 x i8> %res
}
define <16 x i8> @test_masked_i8_to_16_mask3(i8 %s, <16 x i8> %default, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_16_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> %default
  ret <16 x i8> %res
}

define <16 x i8> @test_masked_z_i8_to_16_mask3(i8 %s, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_16_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> zeroinitializer
  ret <16 x i8> %res
}
define <32 x i8> @test_i8_to_32(i8 %s) {
; CHECK-LABEL: test_i8_to_32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastb %edi, %ymm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %res = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  ret <32 x i8> %res
}
define <32 x i8> @test_masked_i8_to_32_mask0(i8 %s, <32 x i8> %default, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_32_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> %default
  ret <32 x i8> %res
}

define <32 x i8> @test_masked_z_i8_to_32_mask0(i8 %s, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_32_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> zeroinitializer
  ret <32 x i8> %res
}
define <32 x i8> @test_masked_i8_to_32_mask1(i8 %s, <32 x i8> %default, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_32_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> %default
  ret <32 x i8> %res
}

define <32 x i8> @test_masked_z_i8_to_32_mask1(i8 %s, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_32_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> zeroinitializer
  ret <32 x i8> %res
}
define <32 x i8> @test_masked_i8_to_32_mask2(i8 %s, <32 x i8> %default, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_32_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> %default
  ret <32 x i8> %res
}

define <32 x i8> @test_masked_z_i8_to_32_mask2(i8 %s, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_32_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> zeroinitializer
  ret <32 x i8> %res
}
define <32 x i8> @test_masked_i8_to_32_mask3(i8 %s, <32 x i8> %default, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_32_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> %default
  ret <32 x i8> %res
}

define <32 x i8> @test_masked_z_i8_to_32_mask3(i8 %s, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_32_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> zeroinitializer
  ret <32 x i8> %res
}
define <64 x i8> @test_i8_to_64(i8 %s) {
; CHECK-LABEL: test_i8_to_64:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastb %edi, %zmm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %res = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  ret <64 x i8> %res
}
define <64 x i8> @test_masked_i8_to_64_mask0(i8 %s, <64 x i8> %default, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_64_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> %default
  ret <64 x i8> %res
}

define <64 x i8> @test_masked_z_i8_to_64_mask0(i8 %s, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_64_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> zeroinitializer
  ret <64 x i8> %res
}
define <64 x i8> @test_masked_i8_to_64_mask1(i8 %s, <64 x i8> %default, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_64_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> %default
  ret <64 x i8> %res
}

define <64 x i8> @test_masked_z_i8_to_64_mask1(i8 %s, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_64_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> zeroinitializer
  ret <64 x i8> %res
}
define <64 x i8> @test_masked_i8_to_64_mask2(i8 %s, <64 x i8> %default, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_64_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> %default
  ret <64 x i8> %res
}

define <64 x i8> @test_masked_z_i8_to_64_mask2(i8 %s, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_64_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> zeroinitializer
  ret <64 x i8> %res
}
define <64 x i8> @test_masked_i8_to_64_mask3(i8 %s, <64 x i8> %default, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_64_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> %default
  ret <64 x i8> %res
}

define <64 x i8> @test_masked_z_i8_to_64_mask3(i8 %s, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_64_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastb %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> zeroinitializer
  ret <64 x i8> %res
}
define <8 x i16> @test_i16_to_8(i16 %s) {
; CHECK-LABEL: test_i16_to_8:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastw %edi, %xmm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %res = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  ret <8 x i16> %res
}
define <8 x i16> @test_masked_i16_to_8_mask0(i16 %s, <8 x i16> %default, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_8_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> %default
  ret <8 x i16> %res
}

define <8 x i16> @test_masked_z_i16_to_8_mask0(i16 %s, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_8_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> zeroinitializer
  ret <8 x i16> %res
}
define <8 x i16> @test_masked_i16_to_8_mask1(i16 %s, <8 x i16> %default, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_8_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> %default
  ret <8 x i16> %res
}

define <8 x i16> @test_masked_z_i16_to_8_mask1(i16 %s, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_8_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> zeroinitializer
  ret <8 x i16> %res
}
define <8 x i16> @test_masked_i16_to_8_mask2(i16 %s, <8 x i16> %default, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_8_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> %default
  ret <8 x i16> %res
}

define <8 x i16> @test_masked_z_i16_to_8_mask2(i16 %s, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_8_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> zeroinitializer
  ret <8 x i16> %res
}
define <8 x i16> @test_masked_i16_to_8_mask3(i16 %s, <8 x i16> %default, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_8_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> %default
  ret <8 x i16> %res
}

define <8 x i16> @test_masked_z_i16_to_8_mask3(i16 %s, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_8_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> zeroinitializer
  ret <8 x i16> %res
}
define <16 x i16> @test_i16_to_16(i16 %s) {
; CHECK-LABEL: test_i16_to_16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastw %edi, %ymm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %res = shufflevector <2 x i16> %vec, <2 x i16> 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>
  ret <16 x i16> %res
}
define <16 x i16> @test_masked_i16_to_16_mask0(i16 %s, <16 x i16> %default, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_16_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> %default
  ret <16 x i16> %res
}

define <16 x i16> @test_masked_z_i16_to_16_mask0(i16 %s, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_16_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> zeroinitializer
  ret <16 x i16> %res
}
define <16 x i16> @test_masked_i16_to_16_mask1(i16 %s, <16 x i16> %default, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_16_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> %default
  ret <16 x i16> %res
}

define <16 x i16> @test_masked_z_i16_to_16_mask1(i16 %s, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_16_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> zeroinitializer
  ret <16 x i16> %res
}
define <16 x i16> @test_masked_i16_to_16_mask2(i16 %s, <16 x i16> %default, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_16_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> %default
  ret <16 x i16> %res
}

define <16 x i16> @test_masked_z_i16_to_16_mask2(i16 %s, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_16_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> zeroinitializer
  ret <16 x i16> %res
}
define <16 x i16> @test_masked_i16_to_16_mask3(i16 %s, <16 x i16> %default, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_16_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> %default
  ret <16 x i16> %res
}

define <16 x i16> @test_masked_z_i16_to_16_mask3(i16 %s, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_16_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> zeroinitializer
  ret <16 x i16> %res
}
define <32 x i16> @test_i16_to_32(i16 %s) {
; CHECK-LABEL: test_i16_to_32:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastw %edi, %zmm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %res = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  ret <32 x i16> %res
}
define <32 x i16> @test_masked_i16_to_32_mask0(i16 %s, <32 x i16> %default, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_32_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> %default
  ret <32 x i16> %res
}

define <32 x i16> @test_masked_z_i16_to_32_mask0(i16 %s, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_32_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> zeroinitializer
  ret <32 x i16> %res
}
define <32 x i16> @test_masked_i16_to_32_mask1(i16 %s, <32 x i16> %default, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_32_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> %default
  ret <32 x i16> %res
}

define <32 x i16> @test_masked_z_i16_to_32_mask1(i16 %s, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_32_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> zeroinitializer
  ret <32 x i16> %res
}
define <32 x i16> @test_masked_i16_to_32_mask2(i16 %s, <32 x i16> %default, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_32_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> %default
  ret <32 x i16> %res
}

define <32 x i16> @test_masked_z_i16_to_32_mask2(i16 %s, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_32_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> zeroinitializer
  ret <32 x i16> %res
}
define <32 x i16> @test_masked_i16_to_32_mask3(i16 %s, <32 x i16> %default, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_32_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> %default
  ret <32 x i16> %res
}

define <32 x i16> @test_masked_z_i16_to_32_mask3(i16 %s, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_32_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastw %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> zeroinitializer
  ret <32 x i16> %res
}
define <4 x i32> @test_i32_to_4(i32 %s) {
; CHECK-LABEL: test_i32_to_4:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastd %edi, %xmm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %res = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  ret <4 x i32> %res
}
define <4 x i32> @test_masked_i32_to_4_mask0(i32 %s, <4 x i32> %default, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_4_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
  ret <4 x i32> %res
}

define <4 x i32> @test_masked_z_i32_to_4_mask0(i32 %s, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_4_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
  ret <4 x i32> %res
}
define <4 x i32> @test_masked_i32_to_4_mask1(i32 %s, <4 x i32> %default, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_4_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
  ret <4 x i32> %res
}

define <4 x i32> @test_masked_z_i32_to_4_mask1(i32 %s, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_4_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
  ret <4 x i32> %res
}
define <4 x i32> @test_masked_i32_to_4_mask2(i32 %s, <4 x i32> %default, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_4_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
  ret <4 x i32> %res
}

define <4 x i32> @test_masked_z_i32_to_4_mask2(i32 %s, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_4_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
  ret <4 x i32> %res
}
define <4 x i32> @test_masked_i32_to_4_mask3(i32 %s, <4 x i32> %default, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_4_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
  ret <4 x i32> %res
}

define <4 x i32> @test_masked_z_i32_to_4_mask3(i32 %s, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_4_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
  ret <4 x i32> %res
}
define <8 x i32> @test_i32_to_8(i32 %s) {
; CHECK-LABEL: test_i32_to_8:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastd %edi, %ymm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %res = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  ret <8 x i32> %res
}
define <8 x i32> @test_masked_i32_to_8_mask0(i32 %s, <8 x i32> %default, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_8_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
  ret <8 x i32> %res
}

define <8 x i32> @test_masked_z_i32_to_8_mask0(i32 %s, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_8_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
  ret <8 x i32> %res
}
define <8 x i32> @test_masked_i32_to_8_mask1(i32 %s, <8 x i32> %default, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_8_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
  ret <8 x i32> %res
}

define <8 x i32> @test_masked_z_i32_to_8_mask1(i32 %s, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_8_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
  ret <8 x i32> %res
}
define <8 x i32> @test_masked_i32_to_8_mask2(i32 %s, <8 x i32> %default, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_8_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
  ret <8 x i32> %res
}

define <8 x i32> @test_masked_z_i32_to_8_mask2(i32 %s, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_8_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
  ret <8 x i32> %res
}
define <8 x i32> @test_masked_i32_to_8_mask3(i32 %s, <8 x i32> %default, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_8_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
  ret <8 x i32> %res
}

define <8 x i32> @test_masked_z_i32_to_8_mask3(i32 %s, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_8_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
  ret <8 x i32> %res
}
define <16 x i32> @test_i32_to_16(i32 %s) {
; CHECK-LABEL: test_i32_to_16:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastd %edi, %zmm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %res = shufflevector <2 x i32> %vec, <2 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>
  ret <16 x i32> %res
}
define <16 x i32> @test_masked_i32_to_16_mask0(i32 %s, <16 x i32> %default, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_16_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
  ret <16 x i32> %res
}

define <16 x i32> @test_masked_z_i32_to_16_mask0(i32 %s, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_16_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
  ret <16 x i32> %res
}
define <16 x i32> @test_masked_i32_to_16_mask1(i32 %s, <16 x i32> %default, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_16_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
  ret <16 x i32> %res
}

define <16 x i32> @test_masked_z_i32_to_16_mask1(i32 %s, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_16_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
  ret <16 x i32> %res
}
define <16 x i32> @test_masked_i32_to_16_mask2(i32 %s, <16 x i32> %default, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_16_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
  ret <16 x i32> %res
}

define <16 x i32> @test_masked_z_i32_to_16_mask2(i32 %s, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_16_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
  ret <16 x i32> %res
}
define <16 x i32> @test_masked_i32_to_16_mask3(i32 %s, <16 x i32> %default, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_16_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
  ret <16 x i32> %res
}

define <16 x i32> @test_masked_z_i32_to_16_mask3(i32 %s, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_16_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastd %edi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
  ret <16 x i32> %res
}
define <2 x i64> @test_i64_to_2(i64 %s) {
; CHECK-LABEL: test_i64_to_2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastq %rdi, %xmm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %res = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  ret <2 x i64> %res
}
define <2 x i64> @test_masked_i64_to_2_mask0(i64 %s, <2 x i64> %default, <2 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_2_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %cmp = icmp eq <2 x i64> %mask, zeroinitializer
  %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> %default
  ret <2 x i64> %res
}

define <2 x i64> @test_masked_z_i64_to_2_mask0(i64 %s, <2 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_2_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %cmp = icmp eq <2 x i64> %mask, zeroinitializer
  %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> zeroinitializer
  ret <2 x i64> %res
}
define <2 x i64> @test_masked_i64_to_2_mask1(i64 %s, <2 x i64> %default, <2 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_2_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %xmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %cmp = icmp eq <2 x i64> %mask, zeroinitializer
  %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> %default
  ret <2 x i64> %res
}

define <2 x i64> @test_masked_z_i64_to_2_mask1(i64 %s, <2 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_2_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %cmp = icmp eq <2 x i64> %mask, zeroinitializer
  %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> zeroinitializer
  ret <2 x i64> %res
}
define <4 x i64> @test_i64_to_4(i64 %s) {
; CHECK-LABEL: test_i64_to_4:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastq %rdi, %ymm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %res = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  ret <4 x i64> %res
}
define <4 x i64> @test_masked_i64_to_4_mask0(i64 %s, <4 x i64> %default, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_4_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
  ret <4 x i64> %res
}

define <4 x i64> @test_masked_z_i64_to_4_mask0(i64 %s, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_4_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
  ret <4 x i64> %res
}
define <4 x i64> @test_masked_i64_to_4_mask1(i64 %s, <4 x i64> %default, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_4_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
  ret <4 x i64> %res
}

define <4 x i64> @test_masked_z_i64_to_4_mask1(i64 %s, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_4_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
  ret <4 x i64> %res
}
define <4 x i64> @test_masked_i64_to_4_mask2(i64 %s, <4 x i64> %default, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_4_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
  ret <4 x i64> %res
}

define <4 x i64> @test_masked_z_i64_to_4_mask2(i64 %s, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_4_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
  ret <4 x i64> %res
}
define <4 x i64> @test_masked_i64_to_4_mask3(i64 %s, <4 x i64> %default, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_4_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %ymm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
  ret <4 x i64> %res
}

define <4 x i64> @test_masked_z_i64_to_4_mask3(i64 %s, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_4_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
  ret <4 x i64> %res
}
define <8 x i64> @test_i64_to_8(i64 %s) {
; CHECK-LABEL: test_i64_to_8:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastq %rdi, %zmm0
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %res = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  ret <8 x i64> %res
}
define <8 x i64> @test_masked_i64_to_8_mask0(i64 %s, <8 x i64> %default, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_8_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
  ret <8 x i64> %res
}

define <8 x i64> @test_masked_z_i64_to_8_mask0(i64 %s, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_8_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
  ret <8 x i64> %res
}
define <8 x i64> @test_masked_i64_to_8_mask1(i64 %s, <8 x i64> %default, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_8_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
  ret <8 x i64> %res
}

define <8 x i64> @test_masked_z_i64_to_8_mask1(i64 %s, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_8_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
  ret <8 x i64> %res
}
define <8 x i64> @test_masked_i64_to_8_mask2(i64 %s, <8 x i64> %default, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_8_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
  ret <8 x i64> %res
}

define <8 x i64> @test_masked_z_i64_to_8_mask2(i64 %s, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_8_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
  ret <8 x i64> %res
}
define <8 x i64> @test_masked_i64_to_8_mask3(i64 %s, <8 x i64> %default, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_8_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %zmm0 {%k1}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
  ret <8 x i64> %res
}

define <8 x i64> @test_masked_z_i64_to_8_mask3(i64 %s, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_8_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastq %rdi, %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
  ret <8 x i64> %res
}
define <16 x i8> @test_i8_to_16_mem(i8* %p) {
; CHECK-LABEL: test_i8_to_16_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastb (%rdi), %xmm0
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %res = shufflevector <2 x i8> %vec, <2 x i8> 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>
  ret <16 x i8> %res
}
define <16 x i8> @test_masked_i8_to_16_mem_mask0(i8* %p, <16 x i8> %default, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_16_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> %default
  ret <16 x i8> %res
}

define <16 x i8> @test_masked_z_i8_to_16_mem_mask0(i8* %p, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_16_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> zeroinitializer
  ret <16 x i8> %res
}
define <16 x i8> @test_masked_i8_to_16_mem_mask1(i8* %p, <16 x i8> %default, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_16_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> %default
  ret <16 x i8> %res
}

define <16 x i8> @test_masked_z_i8_to_16_mem_mask1(i8* %p, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_16_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> zeroinitializer
  ret <16 x i8> %res
}
define <16 x i8> @test_masked_i8_to_16_mem_mask2(i8* %p, <16 x i8> %default, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_16_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> %default
  ret <16 x i8> %res
}

define <16 x i8> @test_masked_z_i8_to_16_mem_mask2(i8* %p, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_16_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> zeroinitializer
  ret <16 x i8> %res
}
define <16 x i8> @test_masked_i8_to_16_mem_mask3(i8* %p, <16 x i8> %default, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_16_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> %default
  ret <16 x i8> %res
}

define <16 x i8> @test_masked_z_i8_to_16_mem_mask3(i8* %p, <16 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_16_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> 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>
  %cmp = icmp eq <16 x i8> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i8> %shuf, <16 x i8> zeroinitializer
  ret <16 x i8> %res
}
define <32 x i8> @test_i8_to_32_mem(i8* %p) {
; CHECK-LABEL: test_i8_to_32_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastb (%rdi), %ymm0
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %res = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  ret <32 x i8> %res
}
define <32 x i8> @test_masked_i8_to_32_mem_mask0(i8* %p, <32 x i8> %default, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_32_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> %default
  ret <32 x i8> %res
}

define <32 x i8> @test_masked_z_i8_to_32_mem_mask0(i8* %p, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_32_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> zeroinitializer
  ret <32 x i8> %res
}
define <32 x i8> @test_masked_i8_to_32_mem_mask1(i8* %p, <32 x i8> %default, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_32_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> %default
  ret <32 x i8> %res
}

define <32 x i8> @test_masked_z_i8_to_32_mem_mask1(i8* %p, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_32_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> zeroinitializer
  ret <32 x i8> %res
}
define <32 x i8> @test_masked_i8_to_32_mem_mask2(i8* %p, <32 x i8> %default, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_32_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> %default
  ret <32 x i8> %res
}

define <32 x i8> @test_masked_z_i8_to_32_mem_mask2(i8* %p, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_32_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> zeroinitializer
  ret <32 x i8> %res
}
define <32 x i8> @test_masked_i8_to_32_mem_mask3(i8* %p, <32 x i8> %default, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_32_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> %default
  ret <32 x i8> %res
}

define <32 x i8> @test_masked_z_i8_to_32_mem_mask3(i8* %p, <32 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_32_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <32 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, 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>
  %cmp = icmp eq <32 x i8> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i8> %shuf, <32 x i8> zeroinitializer
  ret <32 x i8> %res
}
define <64 x i8> @test_i8_to_64_mem(i8* %p) {
; CHECK-LABEL: test_i8_to_64_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastb (%rdi), %zmm0
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %res = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  ret <64 x i8> %res
}
define <64 x i8> @test_masked_i8_to_64_mem_mask0(i8* %p, <64 x i8> %default, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_64_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> %default
  ret <64 x i8> %res
}

define <64 x i8> @test_masked_z_i8_to_64_mem_mask0(i8* %p, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_64_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> zeroinitializer
  ret <64 x i8> %res
}
define <64 x i8> @test_masked_i8_to_64_mem_mask1(i8* %p, <64 x i8> %default, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_64_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> %default
  ret <64 x i8> %res
}

define <64 x i8> @test_masked_z_i8_to_64_mem_mask1(i8* %p, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_64_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> zeroinitializer
  ret <64 x i8> %res
}
define <64 x i8> @test_masked_i8_to_64_mem_mask2(i8* %p, <64 x i8> %default, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_64_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> %default
  ret <64 x i8> %res
}

define <64 x i8> @test_masked_z_i8_to_64_mem_mask2(i8* %p, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_64_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> zeroinitializer
  ret <64 x i8> %res
}
define <64 x i8> @test_masked_i8_to_64_mem_mask3(i8* %p, <64 x i8> %default, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_i8_to_64_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> %default
  ret <64 x i8> %res
}

define <64 x i8> @test_masked_z_i8_to_64_mem_mask3(i8* %p, <64 x i8> %mask) {
; CHECK-LABEL: test_masked_z_i8_to_64_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmb %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastb (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i8, i8* %p
  %vec = insertelement <2 x i8> undef, i8 %s, i32 0
  %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <64 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, 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, 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, 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>
  %cmp = icmp eq <64 x i8> %mask, zeroinitializer
  %res = select <64 x i1> %cmp, <64 x i8> %shuf, <64 x i8> zeroinitializer
  ret <64 x i8> %res
}
define <8 x i16> @test_i16_to_8_mem(i16* %p) {
; CHECK-LABEL: test_i16_to_8_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastw (%rdi), %xmm0
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %res = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  ret <8 x i16> %res
}
define <8 x i16> @test_masked_i16_to_8_mem_mask0(i16* %p, <8 x i16> %default, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_8_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> %default
  ret <8 x i16> %res
}

define <8 x i16> @test_masked_z_i16_to_8_mem_mask0(i16* %p, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_8_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> zeroinitializer
  ret <8 x i16> %res
}
define <8 x i16> @test_masked_i16_to_8_mem_mask1(i16* %p, <8 x i16> %default, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_8_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> %default
  ret <8 x i16> %res
}

define <8 x i16> @test_masked_z_i16_to_8_mem_mask1(i16* %p, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_8_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> zeroinitializer
  ret <8 x i16> %res
}
define <8 x i16> @test_masked_i16_to_8_mem_mask2(i16* %p, <8 x i16> %default, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_8_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> %default
  ret <8 x i16> %res
}

define <8 x i16> @test_masked_z_i16_to_8_mem_mask2(i16* %p, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_8_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> zeroinitializer
  ret <8 x i16> %res
}
define <8 x i16> @test_masked_i16_to_8_mem_mask3(i16* %p, <8 x i16> %default, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_8_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> %default
  ret <8 x i16> %res
}

define <8 x i16> @test_masked_z_i16_to_8_mem_mask3(i16* %p, <8 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_8_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i16> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i16> %shuf, <8 x i16> zeroinitializer
  ret <8 x i16> %res
}
define <16 x i16> @test_i16_to_16_mem(i16* %p) {
; CHECK-LABEL: test_i16_to_16_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastw (%rdi), %ymm0
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %res = shufflevector <2 x i16> %vec, <2 x i16> 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>
  ret <16 x i16> %res
}
define <16 x i16> @test_masked_i16_to_16_mem_mask0(i16* %p, <16 x i16> %default, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_16_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> %default
  ret <16 x i16> %res
}

define <16 x i16> @test_masked_z_i16_to_16_mem_mask0(i16* %p, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_16_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> zeroinitializer
  ret <16 x i16> %res
}
define <16 x i16> @test_masked_i16_to_16_mem_mask1(i16* %p, <16 x i16> %default, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_16_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> %default
  ret <16 x i16> %res
}

define <16 x i16> @test_masked_z_i16_to_16_mem_mask1(i16* %p, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_16_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> zeroinitializer
  ret <16 x i16> %res
}
define <16 x i16> @test_masked_i16_to_16_mem_mask2(i16* %p, <16 x i16> %default, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_16_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> %default
  ret <16 x i16> %res
}

define <16 x i16> @test_masked_z_i16_to_16_mem_mask2(i16* %p, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_16_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> zeroinitializer
  ret <16 x i16> %res
}
define <16 x i16> @test_masked_i16_to_16_mem_mask3(i16* %p, <16 x i16> %default, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_16_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> %default
  ret <16 x i16> %res
}

define <16 x i16> @test_masked_z_i16_to_16_mem_mask3(i16* %p, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_16_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> 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>
  %cmp = icmp eq <16 x i16> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i16> %shuf, <16 x i16> zeroinitializer
  ret <16 x i16> %res
}
define <32 x i16> @test_i16_to_32_mem(i16* %p) {
; CHECK-LABEL: test_i16_to_32_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastw (%rdi), %zmm0
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %res = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  ret <32 x i16> %res
}
define <32 x i16> @test_masked_i16_to_32_mem_mask0(i16* %p, <32 x i16> %default, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_32_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> %default
  ret <32 x i16> %res
}

define <32 x i16> @test_masked_z_i16_to_32_mem_mask0(i16* %p, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_32_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> zeroinitializer
  ret <32 x i16> %res
}
define <32 x i16> @test_masked_i16_to_32_mem_mask1(i16* %p, <32 x i16> %default, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_32_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> %default
  ret <32 x i16> %res
}

define <32 x i16> @test_masked_z_i16_to_32_mem_mask1(i16* %p, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_32_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> zeroinitializer
  ret <32 x i16> %res
}
define <32 x i16> @test_masked_i16_to_32_mem_mask2(i16* %p, <32 x i16> %default, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_32_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> %default
  ret <32 x i16> %res
}

define <32 x i16> @test_masked_z_i16_to_32_mem_mask2(i16* %p, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_32_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> zeroinitializer
  ret <32 x i16> %res
}
define <32 x i16> @test_masked_i16_to_32_mem_mask3(i16* %p, <32 x i16> %default, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_i16_to_32_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> %default
  ret <32 x i16> %res
}

define <32 x i16> @test_masked_z_i16_to_32_mem_mask3(i16* %p, <32 x i16> %mask) {
; CHECK-LABEL: test_masked_z_i16_to_32_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmw %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastw (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i16, i16* %p
  %vec = insertelement <2 x i16> undef, i16 %s, i32 0
  %shuf = shufflevector <2 x i16> %vec, <2 x i16> undef, <32 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, 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>
  %cmp = icmp eq <32 x i16> %mask, zeroinitializer
  %res = select <32 x i1> %cmp, <32 x i16> %shuf, <32 x i16> zeroinitializer
  ret <32 x i16> %res
}
define <4 x i32> @test_i32_to_4_mem(i32* %p) {
; CHECK-LABEL: test_i32_to_4_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vbroadcastss (%rdi), %xmm0
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %res = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  ret <4 x i32> %res
}
define <4 x i32> @test_masked_i32_to_4_mem_mask0(i32* %p, <4 x i32> %default, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_4_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
  ret <4 x i32> %res
}

define <4 x i32> @test_masked_z_i32_to_4_mem_mask0(i32* %p, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_4_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
  ret <4 x i32> %res
}
define <4 x i32> @test_masked_i32_to_4_mem_mask1(i32* %p, <4 x i32> %default, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_4_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
  ret <4 x i32> %res
}

define <4 x i32> @test_masked_z_i32_to_4_mem_mask1(i32* %p, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_4_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
  ret <4 x i32> %res
}
define <4 x i32> @test_masked_i32_to_4_mem_mask2(i32* %p, <4 x i32> %default, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_4_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
  ret <4 x i32> %res
}

define <4 x i32> @test_masked_z_i32_to_4_mem_mask2(i32* %p, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_4_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
  ret <4 x i32> %res
}
define <4 x i32> @test_masked_i32_to_4_mem_mask3(i32* %p, <4 x i32> %default, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_4_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
  ret <4 x i32> %res
}

define <4 x i32> @test_masked_z_i32_to_4_mem_mask3(i32* %p, <4 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_4_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i32> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
  ret <4 x i32> %res
}
define <8 x i32> @test_i32_to_8_mem(i32* %p) {
; CHECK-LABEL: test_i32_to_8_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vbroadcastss (%rdi), %ymm0
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %res = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  ret <8 x i32> %res
}
define <8 x i32> @test_masked_i32_to_8_mem_mask0(i32* %p, <8 x i32> %default, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_8_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
  ret <8 x i32> %res
}

define <8 x i32> @test_masked_z_i32_to_8_mem_mask0(i32* %p, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_8_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
  ret <8 x i32> %res
}
define <8 x i32> @test_masked_i32_to_8_mem_mask1(i32* %p, <8 x i32> %default, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_8_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
  ret <8 x i32> %res
}

define <8 x i32> @test_masked_z_i32_to_8_mem_mask1(i32* %p, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_8_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
  ret <8 x i32> %res
}
define <8 x i32> @test_masked_i32_to_8_mem_mask2(i32* %p, <8 x i32> %default, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_8_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
  ret <8 x i32> %res
}

define <8 x i32> @test_masked_z_i32_to_8_mem_mask2(i32* %p, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_8_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
  ret <8 x i32> %res
}
define <8 x i32> @test_masked_i32_to_8_mem_mask3(i32* %p, <8 x i32> %default, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_8_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
  ret <8 x i32> %res
}

define <8 x i32> @test_masked_z_i32_to_8_mem_mask3(i32* %p, <8 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_8_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i32> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
  ret <8 x i32> %res
}
define <16 x i32> @test_i32_to_16_mem(i32* %p) {
; CHECK-LABEL: test_i32_to_16_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vbroadcastss (%rdi), %zmm0
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %res = shufflevector <2 x i32> %vec, <2 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>
  ret <16 x i32> %res
}
define <16 x i32> @test_masked_i32_to_16_mem_mask0(i32* %p, <16 x i32> %default, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_16_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
  ret <16 x i32> %res
}

define <16 x i32> @test_masked_z_i32_to_16_mem_mask0(i32* %p, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_16_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
  ret <16 x i32> %res
}
define <16 x i32> @test_masked_i32_to_16_mem_mask1(i32* %p, <16 x i32> %default, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_16_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
  ret <16 x i32> %res
}

define <16 x i32> @test_masked_z_i32_to_16_mem_mask1(i32* %p, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_16_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
  ret <16 x i32> %res
}
define <16 x i32> @test_masked_i32_to_16_mem_mask2(i32* %p, <16 x i32> %default, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_16_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
  ret <16 x i32> %res
}

define <16 x i32> @test_masked_z_i32_to_16_mem_mask2(i32* %p, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_16_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
  ret <16 x i32> %res
}
define <16 x i32> @test_masked_i32_to_16_mem_mask3(i32* %p, <16 x i32> %default, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_i32_to_16_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
  ret <16 x i32> %res
}

define <16 x i32> @test_masked_z_i32_to_16_mem_mask3(i32* %p, <16 x i32> %mask) {
; CHECK-LABEL: test_masked_z_i32_to_16_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmd %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastd (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i32, i32* %p
  %vec = insertelement <2 x i32> undef, i32 %s, i32 0
  %shuf = shufflevector <2 x i32> %vec, <2 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>
  %cmp = icmp eq <16 x i32> %mask, zeroinitializer
  %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
  ret <16 x i32> %res
}
define <2 x i64> @test_i64_to_2_mem(i64* %p) {
; CHECK-LABEL: test_i64_to_2_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vpbroadcastq (%rdi), %xmm0
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %res = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  ret <2 x i64> %res
}
define <2 x i64> @test_masked_i64_to_2_mem_mask0(i64* %p, <2 x i64> %default, <2 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_2_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %cmp = icmp eq <2 x i64> %mask, zeroinitializer
  %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> %default
  ret <2 x i64> %res
}

define <2 x i64> @test_masked_z_i64_to_2_mem_mask0(i64* %p, <2 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_2_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %cmp = icmp eq <2 x i64> %mask, zeroinitializer
  %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> zeroinitializer
  ret <2 x i64> %res
}
define <2 x i64> @test_masked_i64_to_2_mem_mask1(i64* %p, <2 x i64> %default, <2 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_2_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %xmm1, %xmm1, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %xmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %cmp = icmp eq <2 x i64> %mask, zeroinitializer
  %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> %default
  ret <2 x i64> %res
}

define <2 x i64> @test_masked_z_i64_to_2_mem_mask1(i64* %p, <2 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_2_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %xmm0, %xmm0, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %xmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
  %cmp = icmp eq <2 x i64> %mask, zeroinitializer
  %res = select <2 x i1> %cmp, <2 x i64> %shuf, <2 x i64> zeroinitializer
  ret <2 x i64> %res
}
define <4 x i64> @test_i64_to_4_mem(i64* %p) {
; CHECK-LABEL: test_i64_to_4_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vbroadcastsd (%rdi), %ymm0
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %res = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  ret <4 x i64> %res
}
define <4 x i64> @test_masked_i64_to_4_mem_mask0(i64* %p, <4 x i64> %default, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_4_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
  ret <4 x i64> %res
}

define <4 x i64> @test_masked_z_i64_to_4_mem_mask0(i64* %p, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_4_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
  ret <4 x i64> %res
}
define <4 x i64> @test_masked_i64_to_4_mem_mask1(i64* %p, <4 x i64> %default, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_4_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
  ret <4 x i64> %res
}

define <4 x i64> @test_masked_z_i64_to_4_mem_mask1(i64* %p, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_4_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
  ret <4 x i64> %res
}
define <4 x i64> @test_masked_i64_to_4_mem_mask2(i64* %p, <4 x i64> %default, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_4_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
  ret <4 x i64> %res
}

define <4 x i64> @test_masked_z_i64_to_4_mem_mask2(i64* %p, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_4_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
  ret <4 x i64> %res
}
define <4 x i64> @test_masked_i64_to_4_mem_mask3(i64* %p, <4 x i64> %default, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_4_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm1, %ymm1, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %ymm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
  ret <4 x i64> %res
}

define <4 x i64> @test_masked_z_i64_to_4_mem_mask3(i64* %p, <4 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_4_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %ymm0, %ymm0, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %ymm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <4 x i64> %mask, zeroinitializer
  %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
  ret <4 x i64> %res
}
define <8 x i64> @test_i64_to_8_mem(i64* %p) {
; CHECK-LABEL: test_i64_to_8_mem:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vbroadcastsd (%rdi), %zmm0
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %res = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  ret <8 x i64> %res
}
define <8 x i64> @test_masked_i64_to_8_mem_mask0(i64* %p, <8 x i64> %default, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_8_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
  ret <8 x i64> %res
}

define <8 x i64> @test_masked_z_i64_to_8_mem_mask0(i64* %p, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_8_mem_mask0:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
  ret <8 x i64> %res
}
define <8 x i64> @test_masked_i64_to_8_mem_mask1(i64* %p, <8 x i64> %default, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_8_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
  ret <8 x i64> %res
}

define <8 x i64> @test_masked_z_i64_to_8_mem_mask1(i64* %p, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_8_mem_mask1:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
  ret <8 x i64> %res
}
define <8 x i64> @test_masked_i64_to_8_mem_mask2(i64* %p, <8 x i64> %default, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_8_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
  ret <8 x i64> %res
}

define <8 x i64> @test_masked_z_i64_to_8_mem_mask2(i64* %p, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_8_mem_mask2:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
  ret <8 x i64> %res
}
define <8 x i64> @test_masked_i64_to_8_mem_mask3(i64* %p, <8 x i64> %default, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_i64_to_8_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm1, %zmm1, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %zmm0 {%k1}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
  ret <8 x i64> %res
}

define <8 x i64> @test_masked_z_i64_to_8_mem_mask3(i64* %p, <8 x i64> %mask) {
; CHECK-LABEL: test_masked_z_i64_to_8_mem_mask3:
; CHECK:       # %bb.0:
; CHECK-NEXT:    vptestnmq %zmm0, %zmm0, %k1
; CHECK-NEXT:    vpbroadcastq (%rdi), %zmm0 {%k1} {z}
; CHECK-NEXT:    retq
  %s = load i64, i64* %p
  %vec = insertelement <2 x i64> undef, i64 %s, i32 0
  %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
  %cmp = icmp eq <8 x i64> %mask, zeroinitializer
  %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
  ret <8 x i64> %res
}